Radiomics analysis of 3D spheroids captures irradiation-induced structural and textural changes

Abstract

Although radiotherapy (RT) is one of the main approaches used in cancer treatment, differences in clinical response among patients have not been eliminated, underscoring the need for improved methods to analyze radiobiological responses. Elucidation of variable responses requires a reliable preclinical model, among which three-dimensional (3D) cell aggregates have emerged as a new gold standard due to their physiological relevance and close resemblance to tumor microenvironments. Non-destructive assessment of the effect of RT or other treatments typically involves brightfield imaging to quantify spheroid volume, diameter, and circularity. However, a radiomics-based analytical framework has the potential to assess a greater range of extracted features from brightfield images beyond volume, diameter and circularity. To contribute to a better understanding of the variability observed in clinical radiation responses and to assess the potential of emerging 3D models in radiobiology, radiation-induced effects were investigated in a 3D culture model of human embryonic kidney (HEK-293) cells. As a proof of concept for the radiomics-based analysis framework, dose-related radiation-induced responses in HEK-293 spheroids were, for the first time, quantified longitudinally using a label-free, live morpho-textural radiomics approach integrating voxel-calculated volume, sphericity, entropy, energy and contrast. The findings showed that HEK-293 spheroid populations were quantitatively monitored using the label-free approach, effectively capturing both early and progressive radiation-induced changes, including subtle volume, sphericity, entropy, energy and contrast variations over time. Workflow robustness and generalizability were further assessed by analyzing brightfield images of dose-related radiation-induced responses in head and neck squamous cell carcinoma-derived FaDu tumor spheroids. Comparable trends were observed in FaDu tumor spheroids, showing the robustness and cross-model applicability of the radiomics workflow. A limitation of the present study is that the direct association between the extracted radiomic features and radiation-induced biological effects was not validated. Therefore, the clinical translation potential of the proposed approach requires further investigation by subsequent experiments integrating biological endpoints. Nevertheless, the present results demonstrated that the brightfield imaging-based radiomics approach could sensitively detect dose-dependent structural and textural changes in HEK-293 and FaDu spheroids following irradiation, offering a low-cost and non-destructive monitoring tool for subsequent radiobiology research.

Introduction

Cancer refers to a group of diseases characterized by the formation of abnormal cells dividing uncontrollably and having the ability to invade and destroy body tissues. This disease, which can spread to any part of the body, is the second most common cause of death worldwide, causing the death of 1 in every 6 people¹. Therefore, finding safe diagnostic and effective treatment methods plays a crucial role in both protecting public health and improving quality of life. Radiation therapy (RT) has been a cornerstone of cancer treatment since the discovery of radium by Marie Curie in 1898 and the early medical use of X-rays². The primary objective of radiation oncology is to inactivate cancer cells by applying higher and more effective doses of radiation to tumor tissue, while simultaneously protecting healthy tissues from dose-related toxicity³. Achieving this balance demands not only advanced delivery technologies but also robust biological model systems that can accurately capture the complex structural and functional responses of tumors to irradiation in preclinical and clinical settings. Despite technological advancements in modern RT^3–7, the efficiency of radiotherapy treatment regimens differs among tumors and patients, yet it is predominantly influenced by the tumor microenvironment (TME). TME is a complex network of biological, biophysical, biomechanical, and biochemical factors affecting the tumor’s ability to survive and respond to treatment. Two-dimensional (2D) cell culture is a conventional, simplistic, and cost-effective method for in vitro cell cultivation; however, such systems cannot fully reflect the physiological conditions of cells in their actual tissue environment because they fail to represent cell-cell and cell-matrix interactions, microarchitecture, and environmental changes of the in vivo TME that are imperative to treatment response^8,9. At the same time, animal models (xenografts) can provide the advantage of recapitulating in vivo tissue organization, yet their mismatched physiology, limited heterogeneity, and lower mutation rates make them less representative of human cancers¹⁰. Accordingly, a mismatch emerges between preclinical findings and clinical outcomes when these models are used to investigate response to anticancer therapies¹¹. Taken together, these limitations have accelerated interest in three-dimensional (3D) cell-culture systems, which offer in vitro constructs by better approximating in vivo cellular structure and multicellular organization^12,13. Within this framework, 3D cell cultures have emerged as particularly promising since they can be fabricated to mimic the microenvironmental determinants of treatment response. Numerous studies have shown that 3D models improve the evaluation of key biological processes, including morphology, proliferation, responsiveness to stimuli, differentiation, drug metabolism, and protein synthesis due to their capability to recapitulate in in vivo-like cellular behavior under in vitro conditions^14–19. As a result, 3D models provide a substantially more realistic platform for studying cancer cell dynamics, reproducing the structural organization of solid tumors, and probing the pathophysiology of human cancers²⁰. 3D cell cultures are widely used for testing drugs and investigating the effect of radiotherapy treatments in oncology. Al-Ramadan et al. showed a dose-dependent sensitivity and spheroid growth analysis based on diameter measurement in BON-1 pancreatic neuroendocrine tumor spheroids within the 0–6 Gy range, observed seven days post-irradiation through the induction of apoptosis¹⁸. Additionally, head and neck squamous cell carcinoma (HNSCC) cells (FaDu, UT-SCC15, UT-SCC45), lung cancer cells (A549), colorectal carcinoma cells (DLD-1, HCT-116), and pancreatic tumor cells (MiaPaCa2, Panc1) were treated with different concentrations of TAE226 (0–1 μm; 1 or 24 h) without or in combination with irradiation (0–6 Gy, X-ray, single dose). The analysis of spheroids demonstrated enhanced cellular radiosensitivity, particularly of HNSCC cells grown in a 3D cell culture model²¹. Recent work has further demonstrated the value of 3D spheroid models by showing that functionalized reporters enable the identification of spatially distinct regions during spheroid growth and in response to radiotherapy, offering a powerful readout for hypoxia- and radiotherapy-related research¹⁷. Moreover, a study using HCT116 and CAL27 tumor spheroids exposed to radiation (0–20 Gy) and hyperthermia has shown distinct treatment responses not observed in 2D cell cultures, highlighting the added value of 3D models for assessing treatment efficacy²². Zanoni et al. irradiated A549 spheroids using different fractionation schedules delivered via a linear accelerator and showed that 3D morphological metrics such as spheroid volume indicated pronounced radiation-induced changes, particularly 25 days after treatment²³. Similarly, Chen et al. provided a mechanistic insight into protective stress-response mechanisms supporting survival of HNSCC spheroids subjected to hyperthermia and radiotherapy, underscoring spheroid volume growth delay and spheroid control probability assays as useful analytical tools for investigating combined therapies²⁴. As described, such studies highlight the growing interest in the use of 3D models supporting in vitro cancer research, particularly focusing on preclinical radiotherapy monitoring. Although 3D models offer numerous advantages for studying radiation responses, their limited feasibility for standardized, live 3D cell analysis represents an experimental challenge, which points to the need for novel approaches that provide reliable, non-invasive, real-time monitoring of radiation-induced effects. Microscopy imaging remains one of the most informative readouts for investigating 3D cell cultures both qualitatively and quantitatively. Today, different approaches (i.e., phase-contrast imaging and stain-based fluorescence imaging via confocal microscopy) are used to acquire the images of 3D cell cultures. Among them, brightfield (BF) imaging is widely used due to its simplicity, accessibility, and suitability for monitoring live 3D cell cultures (i.e., spheroids) without labeling, making it a standard method for investigating morphology and structural architecture and establishing it as an indispensable tool in laboratories worldwide²⁵. BF images can be used to produce quantitative features relevant to diagnosis and treatment assessment. Radiomics formalizes this concept by mining easily acquired medical images with advanced computational analysis, machine learning, and statistical modeling to derive spatial and textural features that encode heterogeneity and other latent biological properties at minimal additional cost^26–28. In radiomics, quantitative features are typically classified into five categories: histogram-based, texture, model-based, transform-based, and shape-based features²⁸. Among these, shape and volume are two features that play a central role in tumor characterization. In clinical settings, accurate shape assessment is essential for surgical planning, while precise volume estimation is crucial for assessing treatment efficiency and guiding clinical decision-making^29–32. Therefore, an accurate estimate of tumor volume plays a key role for patient care. Importantly, although volumetric analysis differs substantially from clinical tumor volume assessment in both biological complexity and imaging scale, volume may still offer a simplified experimental framework for evaluating treatment-associated structural changes at the microscopic level. Reliable data acquisition, however, requires robust and standardized volume-assessment approaches. Despite advances in quantitative image analysis, however, spheroid analysis methodologies still depend on coarse, population-level easily extracted endpoints (i.e., diameter or stain-based volume measurements and clonogenic survival assays). Although informative, such approaches typically rely on fixation, staining, or cell lysis, making them labor-intensive, time-consuming, destructive to the sample, and dependent on specialized expertise and costly instrumentation³³. Accordingly, to extend currently available knowledge and address these limitations, this study introduced a simple and standardized radiomics-inspired analysis using AnaSP 3.0³⁴, an open-source software suite specifically designed to reconstruct 3D geometric descriptors from 2D BF spheroid images and extract morphometric and radiomic features for quantifying radiation-induced spheroid growth kinetics. In AnaSP 3.0, reconstructed 3D geometric descriptors are estimated from single 2D BF projections under the assumption of rotational symmetry around the maximum axis rather than being directly measured from true volumetric image stacks. Accordingly, the reported volume represents a reconstructed 3D volume derived from 2D BF images, while the corresponding sphericity values represent estimated sphericity descriptors obtained from the reconstructed spheroid geometry³⁴. Thus, the presented framework should be interpreted as a microscopy-based radiomics-inspired imaging approach rather than a direct equivalent of clinical radiomics applications.

Using BF microscopy, 3D reconstructed volume (derived from voxel counts), estimated sphericity, and textural metrics based on the Gray-Level Co-Occurrence Matrix (GLCM), namely entropy, energy and contrast^28,35,36 were extracted to evaluate radiation response as a function of time and dose. Entropy, a textural descriptor capturing intra-spheroid intensity heterogeneity^37,38, along with GLCM energy, reflecting the uniformity of pixel pair distributions, and GLCM contrast, quantifying local intensity variations between neighboring pixels^28,35,36,39, was evaluated as a potential preclinical imaging biomarker of radiotherapy response for the first time to further complement morphological characterization. The use of multiscale metrics enabled the probing of radiation-induced morpho-textural alterations occurring during spheroid growth of spheroids and following irradiation. To show the feasibility, non-cancerous human embryonic kidney (HEK-293) spheroids and cancerous FaDu spheroids were used as representative 3D models. Additionally, the fluorescent dye Hoechst was used to facilitate comparison with fluorescence-based measurements and support extracted radiomics features. Growth inhibition was also calculated for HEK-293 and FaDu spheroids.

Results

Radiomics characterization of spheroid responses to radiotherapy

HEK-293 spheroids

A series of quantitative analyses were performed to assess the morphological, gray-level, and growth-kinetics characteristics of HEK-293 spheroids under non-irradiated and irradiated conditions. Following single-fraction exposures of 2 Gy, 4 Gy and 16 Gy, the evolution of spheroid morphology was quantitatively investigated at four imaging time points referring to Days 0, 1, 2, and 5 post-irradiation (hereafter denoted as D0, D1, D2, and D5). The images were processed using AnaSP 3.0 to compute 3D quantitative descriptors, including volume and sphericity, as well as gray-level-derived entropy from the binary mask projections. Moreover, GLCM-based analysis, including energy and contrast measurements, was performed using a custom-written MATLAB code (R2023b, MathWorks, Natick, MA, USA). Using the extracted 3D reconstructed volume metric, the growth-inhibition (GI) parameter was also derived to quantify the radiotherapy effects. D0 corresponds to the day of irradiation. Figure 1A shows representative BF images of the spheroids, starting with D0 and progressing through D5. The middle panel shows corresponding 3D volume reconstructions obtained using AnaSP 3.0^32,34, essentially showing the control spheroids growing and the irradiated spheroids growing more slowly in a dose-responsive manner. Figure 1B presents false-color maps highlighting high-density regions within the spheroids, illustrating changes in structural inhomogeneity. This inhomogeneity may reflect radiation-associated structural alterations potentially related to cell-cycle arrest around D1–D2. The shape of spheroids, with oxygen and nutrient gradients to inner cells, means the inner cells are often hypoxic and radioresistant, so early bulk growth can look similar to controls^40,41. For D2–D5, higher irradiation doses induced increased internal spheroid heterogeneity and structural alterations, which may be associated with radiation-related cellular damage processes such as mitotic catastrophe, apoptotic changes, or necrotic alterations, accompanied by central accumulation of damaged or non-viable regions within the spheroids. The 3D volume reconstructions allowed visualization of the overall spheroid morphology and showed time- and dose-dependent shape differences, as demonstrated in Fig. 1C.

Figure 1.

Multi-modal visualization of non-irradiated and irradiated HEK-293 spheroids at all time points. Each vertical column refers to one dose exposure (Control, 2 Gy, 4 Gy and 16 Gy). (A) Brightfield image of the spheroids, (B) False-color image, showing cell network inhomogeneity. The scale bar is 100 μm, (C) 3D volume reconstruction produced using AnaSP 3.0 and visualized with an HSV colormap, in which hue shows the relative surface depth of the spheroid.

Analysis of the radiomics results shows a significant wealth of biological data. Figure 2A shows normalized growth kinetics, defined as Volume/Volume_D0, over time. Normalization of volumes to the group-specific median volume at D0 minimized variability from inherent biological differences among spheroids, thereby enabling more consistent and reliable comparisons between dose groups. By D5, Control and 2 Gy spheroids continued to grow progressively, as expected due to the ability of cells to recover from this relatively low radiation dose. Conversely, the 4 Gy and 16 Gy dose exposures showed markedly attenuated growth. The pronounced decrease in spheroid volume visualized in the irradiated groups relative to non-irradiated spheroids exhibited a clear dose-dependent suppression of volume expansion (p < 0.01 for Control vs. 16 Gy at D2; p < 0.001 for both 4 Gy and 16 Gy vs. Control at D2 & D5). These findings confirmed that increasing dose resulted in consistently greater suppression of spheroid growth, with 16 Gy showing the most pronounced growth inhibition. From a biological perspective, the growth suppression observed at both 4 Gy and 16 Gy may reflect overlapping radiation-associated responses potentially related to impaired proliferative capacity and radiation-induced cell damage processes, which may contribute to the observed morphological deterioration and volumetric loss across these dose ranges.^33,42,43

Figure 2.

(A) Median values of volume, sphericity, entropy, energy and contrast metrics normalized to the group-specific median values at baseline (D0 = 1) following radiation exposures of Control, 2 Gy, 4 Gy and 16 Gy. (B) Effect sizes for comparisons between Control and each radiation dose group (2 Gy, 4 Gy, and 16 Gy) across all time points (D0–D5) for volume, sphericity, entropy, energy, and contrast. Data are shown as medians of spheroids per treatment arm from n=3 independent experiments. Statistical significance was defined as follows: *P<0.05; **P<0.01; ***P<0.001.

Beyond 3D reconstructed volume quantification, sphericity, and textural metrics based on the GLCM, namely entropy, energy and contrast were computed as complementary metrics to further characterize the responses of spheroids to radiotherapy. Figure 2A also demonstrates the temporal evolution of sphericity, entropy, energy and contrast in HEK-293 spheroids for dose groups from D0 to D5. Entropy analysis revealed significant differences between 4 Gy and 16 Gy irradiated spheroids that were not nearly as evident from the simple volumetric analysis. The complete lack of entropic change for 16 Gy after D0 (entropy values remained at or below baseline throughout the experiment, showing significant relative drops at D1, D2, and D5 (p < 0.001) showed a static, non-expanding central spheroid region, that increased slightly in volume as it began to unravel. The 4 Gy-treated spheroids were expected to display delayed death, with partial shrinkage but heterogenous architecture leading to initial increase in entropy up to D2. Suppressed growth put the spheroids on a reduced entropic path subsequently and by D5, 4 Gy showed a significantly decreased entropy relative to Control (p < 0.001).

For sphericity (Fig. 2A), the control, 2 Gy and 4 Gy groups displayed a modest increase from D0 to D1, while 16 Gy showed a dose-dependent decrease. The greatest decrease in sphericity was observed in 16 Gy, reaching below-baseline values at D1, D2 and D5 (p < 0.001), likely associated with progressive structural disorganization within the spheroid center. Moreover, sphericity was significantly different from Control for the 4 Gy dose at D1 (p = 0.009) and D2 (p = 0.002); however, by D5 these differences were no longer significant, with sphericity values overlapping with Control. In spite of being a low dose, the 2 Gy group showed a statistically distinguishable decrease in sphericity at D2 relative to the Control by (p < 0.001), followed by a significant increase at D5 (p < 0.01), consistent with transient radiation-induced perturbations in cell proliferation and packing geometry early after irradiation, followed through recovery and more cohesive regrowth by D5^33,44,45.

No statistically significant differences were observed among the Control and 2 Gy groups up to D5 for GLCM energy; however, the 4 Gy group showed a significant difference relative to Control at D5 (p < 0.01) (Fig. 2A). In contrast, the 16 Gy group diverged significantly from Control as early as D1 (p < 0.001), which remained significant at subsequent time points. After D2, GLCM energy increased in association with elevated spheroid homogeneity, reflecting more uniform and repetitive spatial patterns³⁹, while entropy reduced as structural randomness within the spheroid diminished, consistent with a non-viable structurally disorganized central region at 16 Gy. Accordingly, radiation-induced loss of cellular heterogeneity decreased image noise, leading to a reciprocal trend between entropy and energy, consistent with prior radiomics findings^46,47.

As shown in Fig. 2A, GLCM contrast exhibited a time- and dose-dependent behavior across treatment groups. In the 2 Gy group, contrast increased significantly at D1 (p = 0.014), was not different from Control at D2 (p = 0.662), and then diverged again at D5, where differences reached significance (p < 0.01). In the 4 Gy group, GLCM contrast increased significantly at D1 (p < 0.01) and D2 (p < 0.001), but subsequently exhibited a declining trend toward D5, at which point the differences were no longer statistically significant. The 16 Gy group displayed strong and sustained decreases at D1, D2 and D5 (p < 0.001), reflecting a more persistent high-dose effect associated with progressive homogenization within the spheroid due to non-viable cells, as evidenced by lower entropy and higher GLCM energy. The GLCM contrast results suggested a treatment-induced reduction in spheroid heterogeneity and were consistent with previously reported results^48,49.

Control-normalized GI ranked the responses as , in which the strongest suppression was observed for 16 Gy. 2 Gy did not prevent spheroid volume expansion, in contrast to 4 Gy and 16 Gy, both of which indicated strong growth-inhibitory effects relative to Control. In accordance with GI, it was minimum for 2 Gy and substantial for 4 Gy with approximately 60% and exceeded 60% for 16 Gy, confirming shrinking rather than stasis. Non-irradiated HEK-293 spheroids continued to expand over time, while ablative doses (4 Gy and 16 Gy) suppressed 3D reconstructed volume growth.

Figure 2B illustrates the effect sizes quantifying the magnitude and direction of differences between Control and each radiation dose group (2 Gy, 4 Gy, and 16 Gy) from D0 to D5 for volume, sphericity, entropy, energy, and contrast.

Effect sizes for all metrics were moderate at D1 and D2 and attenuated by D5 for the Control and 4 Gy comparison. Volume indicated small positive effects mainly at D1, which decreased by D2. Entropy showed a gradual increase toward D5, whereas sphericity demonstrated positive effects at early time points but declined at D5. In parallel, GLCM energy displayed predominantly negative effects at D5, confirming greater spatial uniformity in the 4 Gy groups, while GLCM contrast showed positive effects at all time points.

A positive effect was visualized for volume at D2 and D5, confirming that Control HEK-293 spheroids were distinguishably larger than those exposed to 16 Gy. Entropy showed continuously positive effect sizes from D0 to D5, confirming reduced textural heterogeneity as a function of increased dose at later time points. A negative effect was observed for sphericity, reflecting lower sphericity in the 16 Gy group across the time course. In parallel, GLCM energy showed predominantly negative effect sizes, which demonstrate greater spatial uniformity in the 16 Gy group relative to Control, while GLCM contrast indicated positive effect sizes at D2 and D5, in line with intensity differences caused by radiation-induced structural remodeling.

According to the effect size mapping using Cliff’s δ in Fig. 2B, clear dose-dependent responses were observed when comparing each treated group (2 Gy, 4 Gy, and 16 Gy) to the Control across all metrics. The heatmap showed clear dose-dependent responses, with effect sizes generally increasing in magnitude from 4 Gy to 16 Gy. Both volume and sphericity showed progressively stronger positive effect sizes at 4 Gy and 16 Gy doses, consistent with radiation-induced growth suppression and structural deformation in spheroids. By contrast, effect sizes for 2 Gy volumes largely overlapped with Control across time points, indicating minimal divergence from baseline. Among texture features, GLCM entropy indicated increasingly positive effect sizes with increasing dose, indicating reduced heterogeneity, whereas GLCM energy exhibited increasingly negative effect sizes, reflecting a trend toward greater internal uniformity, which was most pronounced for 16 Gy. In the 4 Gy group, the negative energy effect emerged at D5 and was weaker in magnitude compared with 16 Gy, consistent with a delayed and less severe structural homogenization. Entropy changes in the 2 Gy group closely followed those of Control all time points, confirming limited radiation-induced changes; however, the 2 Gy group still reflected time- and dose-dependent changes in GLCM energy, although the magnitude of these effects was more modest in comparison to 4 Gy and 16 Gy groups. Responses to RT were associated with increased effect sizes for GLCM contrast, with the most pronounced effect visualized at relatively high-dose of 16 Gy, pointing to distinguishable spatial heterogeneity. Taken together, the effect size map revealed a coherent dose-response relationship for both morphological and texture-based metrics, underscoring the sensitivity to radiation-induced morpho-textural responses of HEK-293 spheroids.

Representative stained nuclei images of HEK-293 spheroids were given in Fig. 3 to support the quantitative radiomics findings. In accordance with the total intensity analysis, significant differences at D2 and D5 between the Control, 4 Gy and 16 Gy groups (p < 0.001) were observed . Our analyses were based on BF images because quantitative gray-level and shape metrics derived from Hoechst images can be affected by staining state and depth-dependent signal attenuation in larger spheroid (roughly 500–600 μm)²³. In the present study, median Feret diameters spanned approximately 160–230 μm across all doses and time points, with clear dose-dependent differences at D2 and D5. The number of analyzed spheroids per treatment arm was: D2 (, , , ); D5 (, , , ).

Figure 3.

(A) Fluorescence imaging of Control and irradiated HEK-293 spheroids reveals suppressed growth kinetics 2 and 5 days after treatment. (B) Fluorescence total intensity data for the Control, 2 Gy, 4 Gy and 16 Gy groups. Irradiated spheroids had significantly reduced intensity compared to Controls (p < 0.001), indicating suppressed growth and decreased nuclear signal post-treatment. The scale bar is 100 μm. Data show values. The number of analyzed spheroids per treatment arm was: D2 (, , , ); D5 (, , , ). Statistical significance was defined as follows: *P<0.05; **P<0.01; ***P<0.001.

FaDu spheroids

For FaDu spheroids, a similar analysis pipeline was employed to quantify the evolution of spheroid morphology following single-fraction X-ray exposures of 5 Gy, 7.5 Gy, and 20 Gy, alongside a non-irradiated control group. Due to the primary objective being to delineate dose-dependent structural responses for the FaDu model, the analysis focused on the intrinsic descriptors including volume, sphericity, entropy, energy, and contrast to characterize radiation-induced morphological and textural changes relative to control. To capture short-range and early post-treatment responses, imaging was performed at four time points corresponding to Days 0, 3, 5, and 7 post-irradiation (hereafter referred to as D0, D3, D5, and D7). Panels A-C show the BF images, AnaSP-based 3D volume reconstructions, and false-color maps, respectively, at all time points (Fig. 4).

Figure 4.

Multi-modal visualization of non-irradiated and irradiated FaDu spheroids at all time points. Each vertical column refers to one dose exposure (Control, 5 Gy, 7.5 Gy and 20 Gy). (A) Brightfield image of the spheroids, (B) False-color image, showing cell network inhomogeneity. The scale bar is 100 μm, (C) 3D volume reconstruction produced using AnaSP 3.0 and visualized with an HSV colormap, in which hue shows the relative surface depth of the spheroid.

Normalized growth kinetics were analyzed as Volume/Volume_D0 from D0 to D7, as shown in Fig. 5A. By D3, non-irradiated FaDu spheroids continued to expand in volume, while those treated by 5 Gy, 7.5 Gy and 20 Gy indicated growth suppression at which the divergence became statistically distinguishable, with both 7.5 Gy and 20 Gy showing markedly reduced volume relative to Control (p < 0.001). At later time points, these differences remained evident for 7.5 Gy and 20 Gy. By D5, 7.5 Gy and 20 Gy dose groups continued to decline relative to Control, showing dose-dependent growth arrest (p < 0.001 for both 7.5 Gy and 20 Gy). FaDu spheroids showed a dose-dependent impairment; although 5 Gy spheroids continued to proliferate, their growth significantly restrained relative to Control from D5 onward (p < 0.01), with this difference maintained through D7. In contrast, 7.5 Gy and 20 Gy spheroids displayed sustained growth inhibition, and 20 Gy resulted in the most pronounced anti-proliferative effect in volume expansion by D3. Consistent with the HEK-293 findings, FaDu spheroids also exhibited dose-dependent growth inhibition when assessed by volume. As given for HEK-293 spheroids, cell-death pathways likely account for the decreased proliferation and accompanying volumetric variations in FaDu spheroids after treatment.^33,42,43.

Figure 5.

(A) Median values of volume, sphericity, entropy, energy and contrast metrics normalized to the group-specific median values at baseline (D0 = 1) following radiation exposures of Control, 5 Gy, 7.5 Gy and 20 Gy. (B) Effect sizes for comparisons between Control and each radiation dose group (5 Gy, 7.5 Gy, and 20 Gy) across all time points (D0–D7) for volume, sphericity, entropy, energy, and contrast. Data are shown as medians of spheroids per treatment arm from n=3 independent experiments. Statistical significance was defined as follows: *P<0.05; **P<0.01; ***P<0.001.

For entropy, as shown in Fig. 5A, Control group showed a progressive increase from D0 onward, which was consistent with ongoing growth and increasing internal heterogeneity. While entropy in the 5 Gy group still increased over time, this increase was induced by radiation, contributing to significant differences relative to Control from D5 onward (p < 0.01). Conversely, both 7.5 Gy and 20 Gy gave rise to distinguishable reductions in entropy initiating at D3, even though both were statistically significant at later time points (p < 0.001), the magnitude of the reduction was greater for 20 Gy.

All dose groups showed decreased sphericity at D3, but the impact differed by dose. Statistical comparisons showed that non-irradiated FaDu spheroids distinguished significantly from both 7.5 Gy and 20 Gy (p < 0.001), while the 5 Gy group did not differ from Control at earlier time points but became significantly distinct at D7 (p = 0.003). By D5, sphericity was significantly greater in the 7.5 Gy and 20 Gy groups compared with Control (p < 0.001) and remained unchanged between D5 and D7, demonstrating a partial morphological recovery in the presence of continued growth inhibition. Even though a transient reduction in sphericity was visualized at D3 in all groups, only the higher dose groups (7.5 Gy and 20 Gy) subsequently showed a distinguishable increase in sphericity from D3 to D5, while Control and 5 Gy spheroid did not. This divergence suggested that the rise was not a natural growth-related phenomenon yet rather a radiation-induced structural reorganization. This increase likely reflected peripheral growth arrest and compaction of the inner core, resulting in a smoother and more spherical morphology^33,44,45,50.

GLCM energy values in FaDu spheroids did not differ significantly between the Control and 5 Gy groups across all time points (Fig. 5A). In contrast, both the 7.5 Gy and 20 Gy groups diverged significantly from Control beginning at D3 (p < 0.05 for 7.5 Gy; p < 0.001 for 20 Gy), which remained significant at subsequent time points (p < 0.001 for both 7.5 Gy and 20 Gy). In HEK-293 spheroids, entropy decreased after D2 at both 4 Gy and 16 Gy. Conversely, FaDu spheroids showed a more complex response: entropy increased up to D3, followed through a dose-dependent reduction between D3 and D5, particularly for 7.5 Gy and 20 Gy, and then stabilized at baseline. While both cell lines initially indicated a dose-dependent increase in GLCM energy, FaDu spheroids exhibited a subsequent reduction in energy after D5. This divergence at D5 may suggest a transition from a more uniform microstructure to a re-emerging spatial heterogeneity, potentially driven by radiation-induced differential cell loss and the emergence of patchy viable regions with high and low cell density (i.e., structurally diverse quiescent zone), together with irregular necrotic or fibrotic regions^43,50–52.

Control and 5 Gy groups indicated a progressive increase over time; however, the 5 Gy group diverged significantly from Control by D7 (p < 0.05), with lower contrast values, as shown in Fig. 5A. At higher doses, distinct dose-dependent behaviors were observed: in the 7.5 Gy groups, contrast elevated up to D3, decreased between D3 and D5, becoming significantly different from Control (p < 0.001), and this difference remained through D7. Contrast reduced from D0 to D5, falling below the baseline level, and partially recovered at D7, yet remained significantly lower than Control (p < 0.001). The 20 Gy contrast profile in FaDu spheroids was consistent with HEK-293 spheroids at matched time points (D0-D5). In particular, the change in trend at D5 may show a shift toward spatial heterogeneity, induced by differential radiation responses across spheroids, which could also explain the increase in GLCM contrast, consistent with previous findings⁵¹. In a manner analogous to that observed in HEK-293 spheroids, GLCM contrast and energy showed a reciprocal behavior over time. Moreover, the behavior of GLCM contrast differed from that observed in non-cancerous HEK-293 spheroids, likely reflecting the cancerous phenotype of FaDu spheroids and its effects on texture responses to radiotherapy⁵³.

Control-normalized GI ranked the responses as , in which the strongest suppression was observed for 20 Gy. 5 Gy, 7.5 Gy and 20 Gy all indicated strong growth-inhibitory effects relative to the control. In accordance with GI, it was found 50% for 5 Gy and substantial for 7.5 Gy with approximately 85% and 87% for 20 Gy, confirming shrinking rather than stasis. Non-irradiated FaDu spheroids continued to expand over time, while ablative doses suppressed 3D reconstructed volume growth.

According to the effect size mapping using Cliff’s δ in Fig. 5B, clear dose-dependent responses were observed when comparing each treated group (5 Gy, 7.5 Gy, and 20 Gy) to the Control across all metrics. The heatmap revealed clear dose-dependent responses, with effect sizes generally increasing in magnitude from 5 Gy to 20 Gy. Volumetric analysis showed progressively stronger positive effect sizes at all doses, while sphericity revealed negative effect sizes at 7.5 Gy and 20 Gy, consistent with radiation-induced growth suppression and structural deformation in spheroids. Among texture features, GLCM entropy displayed increasingly positive effect sizes with dose, demonstrating decreased heterogeneity, while energy showed increasingly negative effect sizes, reflecting a trend toward greater internal uniformity. By contrast, responses to RT were associated with increased effect sizes for GLCM contrast, with the most pronounced effect visualized at 20 Gy, suggesting marked spatial heterogeneity. Taken together, the effect size map revealed a coherent dose-response relationship for both morphological and texture-based metrics, underscoring the sensitivity to radiation-induced morpho-textural responses.

Discussion

This study presented, for the first time, the use of label-free, temporal, and longitudinal radiomics for the evaluation of the effect of radiotherapy on HEK-293 and FaDu spheroids. As a proof of concept and to establish a reliable micro-tissue model, HEK-293 spheroids were selected since conventional 2D cell cultures lack an advanced extracellular matrix (ECM), such as that found in cancer cells, meaning they cannot adequately mimic the TME. In contrast, 3D spheroid models bridge the gap between monolayer cell cultures and animal models and better recapitulate in vivo conditions such as communication between cancer cells and their surrounding components^{8,9,20,54–58}. The reason for evaluating RT influences in HEK-293 spheroids was that this model has been widely used in biomedical research, not only as a robust tool for rapid growth and recombinant protein expression due to its human origin⁵⁹ but also as a model with tumorigenic potential⁶⁰. Furthermore, survival curves obtained in radiotherapy experiments either with 3D spheroids or with solid tumors were similar; thus, spheroids were considered to be appropriate in vitro models to mimic basic biological properties of tumor cells in vivo^23,61–63. HEK-293 cells also demonstrated an exceptional ability to self-assemble into compact spheroids without the need for any support of exogenous agents⁶⁴. To broaden the applicability of the radiomics analysis, FaDu spheroids were analyzed as a clinically relevant cancer model with reproducible benchmarks for validating analytical pipelines as a well-characterized HNSCC model. Hence, evaluating radiation-induced responses in both non-cancerous and cancerous spheroids allowed investigation of the generalizability of the approach across distinct biological surroundings and determined whether volume and entropy showed similar decay trends, serving as reliable biomarkers across both models.

BF images of both HEK-293 and FADU spheroids highlighted distinct early and late responses to irradiation via multi-modal visualizations that demonstrated the ability of this framework to visually capture responses to radiation across different spheroid models. Visually, spheroids appear as translucent balls with a well-defined boundary and darker center when imaged through BF microscopy^65,66. Quantitative characterization of spheroid morphology is one key to understanding growth kinetics and treatment response⁶⁷. While direct structural measurements such as width and diameter allow volume estimation, the radiomics approach employed in this study provided a more comprehensive description for both non-cancerous HEK-293 and cancerous FaDu spheroids. Here we included voxel-based 3D reconstructed volume, sphericity, GLCM entropy, GLCM energy and GLCM contrast notably for spheroids deviating from perfect symmetry. Volume proved to be the most sensitive indicator of radiation-dose response, in agreement with previous findings^17,24,68,69. Complementarily, entropy, energy and contrast quantified variations in the heterogeneity of intensity within the spheroid volume, representing microstructural reorganization, while sphericity measured changes in the shape of the spheroid caused by RT. This enabled clear differentiation of the 4 Gy and 16 Gy dosages, which were not as apparent by volumetric analysis. Sphericity teased out subtle differences in the growth patterns of the control and low-dose (2 Gy) spheroids. Together, the five parameters reported different aspects of cell behavior, reinforcing the importance of using multiscale quantitative metrics to investigate their responses to stimuli.

For HEK-293 spheroids, all metrics were normalized to their respective D0 values to minimize baseline and sampling variability. The use of median values of volume were robust to outliers and reduced variability, correcting for inevitable between-group differences in starting metrics arising from spheroid production and handling. Hence, under this scaling, a value of 1 at D0 was used for all experiment groups (Fig. 2A) to show robust growth kinetics rather than initiation offsets. This normalization assumed any early radiotherapy influence did not change radiomics-based metrics at D0. Volume-based growth analysis revealed HEK-293 spheroids to continue to grow from D0 to D5, with a rather linear expansion of the volume distribution through D5, a pattern matching prior observations in uninhibited growth of 3D HEK-293 spheroids^68,70 (Fig. 2A and B).

The first observable effect for HEK-293 spheroids in volume-based growth kinetics emerged at D1, with the 16 Gy dose significantly distinguishable from Control (p < 0.01)—4 Gy dose at D1 showed p = 0.148. By D2, both the 4 Gy and 16 Gy doses showed divergence, continuing through D5 for both treatments. The 4 Gy dose group showed slightly greater volume ratios than 16 Gy. The volumes for both 4 Gy and 16 Gy between D2 and D5 plateaued, representing a sustained but stabilized growth inhibition. Irradiating HEK-293 spheroids with 2 Gy showed no significant growth difference to Control at any time point, showing that growth kinetics based on volume were sustained at low-dose exposure. This pattern also highlighted that radiation-induced growth inhibition was both cumulative and sustained, with higher doses (4 Gy and 16 Gy) showing the most pronounced suppression effect relative to non-irradiated spheroids. Dose escalation to 16 Gy did not lead to a significantly distinguishable growth inhibition compared to 4 Gy for HEK-293 spheroids, suggesting 4 Gy is also a highly effective treatment. This observation was found consistent with previous spheroid-based RT studies reporting plateauing growth inhibition at moderate-to-high single fraction doses⁷¹. Consistently, this pattern was also captured by the derived GI analysis, in which the normalized growth inhibition index was related to radiation dose, showing a GI that was minimum for 2 Gy and substantial for 4 Gy with approximately 60% and exceeded 60% for 16 Gy, confirming inhibition of volume expansion at D5.

As for FaDu spheroids, a similar pattern was observed consistent with the results of HEK-293 spheroids. By normalizing all FaDu measurements to the baseline time point (D0), this study focused on within-dose longitudinal trajectories, allowing direct comparisons between doses at follow-up days (D0, D3, D5 and D7). Non-irradiated FaDu spheroids and those exposed to 5 Gy showed continued volumetric growth over time, with a significant decrease from control emerging at D5 for 5 Gy, while 7.5 Gy and 20 Gy led to a gradual decline in spheroid volume between D3 and D7, with the greatest inhibition observed at 20 Gy (p < 0.001) throughout this interval. This dose-dependent trend was consistent with previous findings in a LNCaP cancer spheroid model exposed to 20 Gy⁷¹. These data confirmed that relative growth kinetics in cancer spheroids could effectively differentiate dose-dependent RT effects.

Volume-based results were further supported through GLCM entropy, GLCM energy and GLCM contrast and sphericity analyses of both HEK293 and FaDu spheroids. In the Control and 2 Gy groups of HEK-293 spheroids, entropy increased with time, consistent with continuous spheroid volume expansion and the development of structural heterogeneity resulting from ongoing cell proliferation. Conversely, entropy showed a steady trend in the 16 Gy group and elevated in the 4 Gy group up to D2; yet, from D2 to D5 both dose groups indicated a clear decline in entropy relative to control. This can be explained by the radiation-induced microstructural homogenization, likely associated with reduced cellular complexity following irradiation, supporting entropy as a sensitive descriptor of spheroid response to RT. This was in line with previous studies, stating that the spheroid entropy was decreased in glioblastoma (U87) spheroids exposed to graphene quantum dots photothermal therapy⁷². In parallel, a concomitant decrease in fluorescence intensity of Hoechst-stained nuclei was observed in irradiated HEK-293 spheroids as shown in Fig. 3, associating with volume quantifications. Spheroids exposed to 4 Gy and 16 Gy showed lower nuclear fluorescence intensity than non-irradiated spheroids at both D2 and D5, indicative of radiation-induced lack of mitosis. Such dose-dependent divergence aligns with previously reported in vitro results, in which higher radiation exposure caused a decrease in cellular proliferation activity, thereby leading to a loss of integrity in 3D culture systems^71,73.

For HEK-293 spheroids, despite being stable in entropy up to D2 in the 16 Gy group, GLCM energy had increasing tendencies over time since it is a measure of homogeneity of an image³⁹. A higher value for this feature showed that the intensity in a spheroid image varied less. Furthermore, the energy in 4 Gy group diverged significantly from Control at D5 as entropy declined. These findings implied that radiation effects primarily changed local microstructural organization rather than global heterogeneity within HEK-293 spheroids. GLCM energy values showed an inverse temporal pattern to entropy, consistent with its mathematical relationship to texture regularity³⁹. The findings align with previous radiomics studies reporting that radiation-induced loss of cellular heterogeneity produced a reciprocal relationship between entropy and energy^46,47. Moreover, GLCM contrast showed time and dose-dependent variations following irradiation (Fig. 2A). Control, 2 Gy and 4 Gy produced transient early increases followed by a gradual decline, while 16 Gy had decreasing tendencies over time, with significantly lower values at D2 (p < 0.01) and D5 (p < 0.001) relative to Control. Since lower GLCM contrast shows reduced spatial variation within spheroids, the more pronounced and sustained decrease in contrast obtained in the 16 Gy group compared with other doses likely demonstrated greater homogeneity, which could be attributed to radiation-induced cell death. The 4 Gy treated spheroids likely underwent mixed senescence and delayed cell death, resulting in partial shrinkage but a more heterogeneous microstructure that emerged as an initial rise in GLCM contrast up to D2. Subsequently, as cellular viability declined, the 4 Gy treated spheroids shifted toward a lower-contrast trajectory by D5, becoming comparable to Control. This did not preclude radiation effects; rather, it suggested that local intensity differences reached similar levels in both groups by D5, potentially due to growth-related remodeling in non-irradiated spheroids and heterogeneous dose responses. This interpretation was further supported by concomitant decreases in entropy and increases in GLCM energy for D2-D5 time interval. Overall, these findings indicate a radiation-induced reduction in spheroid heterogeneity, in line with previous reports^48,49.

Regarding FaDu spheroids, GLCM-based metrics (entropy, energy and contrast) offered converging evidence of radiation-induced variations in intra-spheroid heterogeneity, despite differences in biological background and experimental settings. Particularly, even though the temporal frameworks were not same (HEK-293: D0-D5; FaDu:D0-D7), intermediate phases (D3-D5/7) showed comparable trajectories across both spheroid models, allowing meaningful cross-model evaluation. For FaDu spheroids, entropy changed in parallel with volumetric growth kinetics. Increasing volume was accompanied by increasing entropy, consistent with the progressive structural heterogeneity caused by ongoing cell proliferation. Moreover, higher doses (7.5 Gy and 20 Gy) were also associated with reductions in both volume and entropy. Moreover, GLCM energy differentiated 7.5 Gy and 20 Gy dose responses from Control starting at D3, while  5 Gy remained comparable to Control across all time points, showing a clear dose threshold for observable textural regularization (Fig. 5A). Unlike HEK-293 spheroids, FaDu resulted in a biphasic response, an inital entropy increase followed through a dose-dependent decline and stabilizing thereafter pointing to a more dynamic remodeling process after irradiation by 7.5 Gy and 20 Gy. Despite both cell lines displaying early dose-dependent increases in energy, the subsequent decrease in energy in FaDu spheroids after D5 implied a reconstitution of spatial heterogeneity that may be driven through differential radiation-induced cell loss and the development of patchy viable cell regions interspersed by irregular necrotic and fibrotic areas. With respect to GLCM contrast, a time- and dose-dependent trend was evident. Contrast deviated only mildly from baseline from D0 to D5; however the reversal at D5 may point to a re-emergence of spatial heterogeneity due to variable radiation responses across spheroids⁵¹. As in HEK-293, contrast and energy behaved reciprocally, yet the FaDu-specific patterns highlighted the influence of the malignant phenotype on radiotherapy-induced texture changes⁵³.

Sphericity analysis showed a clear and dose-dependent temporal response in both spheroid models. In HEK-293 spheroids, sphericity demonstrated a transient behavior, characterized by an initial increase at D1 for Control and 2 Gy, followed by a gradual decrease at later time points, pointing to an early physical disruption of spheroid due to a disorganized growth pattern induced by irradiation⁴¹. This finding demonstrated that sphericity was a discriminative metric capable of differentiating Control from 2 Gy treated spheroids. Sphericity decreased steadily with escalating radiation doses, with both the 4 Gy and 16 Gy groups showing reductions over time. The 16 Gy group had the most evident and maintained reduction in sphericity. The dose-dependent decrease in sphericity signified radiation-induced physical disruption in spheroid structure. The present study showed that at elevated dosages, sphericity can capture treatment-induced morphological changes. In contrast to prior studies reporting that sphericity is not dependent on spherical damage following external stimuli^74,75, these findings indicated that sphericity is sensitive to physical disruption of spheroids exposed to high-dose irradiation. Biologically, the observed variations between treated and untreated spheroids may be attributed to decreased cell-cell interactions and loss of adherens junctions⁷⁶.

The sphericity of FaDu spheroids in the 20 Gy showed a significant reduction, by a similar trend observed in the 7.5 Gy group, pointing to an early physical disruption of spheroid due to uncontrolled growth induced by irradiation. This transient decrease can be attributed to boundary irregularities due to spheroid outgrowth and the presence of cellular debris after RT exposure. The variability in sphericity values across dose groups was lost at D7. This trend presumably indicated compaction and reorganization of the leftover viable cell population into a more compact volume⁷⁴. As for sphericity, compared to HEK-293 spheroids, the temporal shift may stem from cell-line-specific dynamics and radiosensitivity²¹, as well as the delayed initiation of monitoring in FaDu. Yet, sphericity proved to be responsive to radiation-induced physical disruption in both spheroid models.

Sphericity decreased across all treatment groups in the early phase following irradiation, showing an initial, global perturbation of FaDu spheroid morphology. This transient decrease can be attributed to boundary irregularities due to spheroid outgrowth and the presence of cellular debris after RT exposure. At later time points, distinct dose-dependent trajectories emerged. In the 7.5 Gy and 20 Gy groups, sphericity recovered and exceeded baseline at later time points, whereas Control and 5 Gy spheroids remained unchanged before progressively decreasing. These distinct responses suggested that sphericity is sensitive to radiation-induced morphological changes and may facilitate discrimination between low and high-dose responses despite similar early behavior.

Together, these findings suggest that, reconstructed 3D volume and estimated sphericity reflected spheroid growth and morphologic dynamics in both cancerous and non-cancerous spheroid models, while GLCM-derived texture descriptors offered complementary information regarding treatment-induced structural heterogeneity and morphologic alterations that were not fully captured by geometric measurements alone, thereby providing a more comprehensive characterization of radiotherapy-induced responses.

The described study has limitations. Firstly, it is limited to spheroids and does not explore radiomics analysis of either 2D monolayers or patient-derived organoid models, and there are intrinsic biological variability between spheroids and organoids^18,77,78. Secondly, only single fraction doses were tested; fractionation and dose-rate influences were not evaluated. The fluorescent study did not begin to assess radiomics analysis of fluorescence imaging, limiting the analysis to BF images. Moreover, confirmation of volume quantification by biological endpoints, such as caspase activation for apoptosis and hypoxia reporters, could further help attribute radiomics analysis to biologically defined pathways. In this regard, although only limited volumetric and radiomics alterations were observed in the 2 Gy group, which represents a conventional clinical fraction dose, these findings should not be interpreted as indicating biological insignificance or the absence of radiation-induced effects. Since complementary assays evaluating DNA damage, viability, proliferation, clonogenic survival, or cell-cycle alterations were not performed, subtle or sublethal radiation-associated responses may not have been fully detected through the quantified metrics in the present study. Accordingly, the metrics may be more sensitive to pronounced higher-dose structural alterations than to early low-dose radiation effects. Together, such refinements would position live, label-free radiomics analysis as a scalable and biologically representative model for investigating radiotherapy effects, advancing studies based on 3D models and bridging the gap between 2D assays and in vivo models. It should be noted that this study solely focused on the methodological establishment of radiomics-based feature extraction from irradiated spheroid images, rather than investigating the specific biological mechanisms underlying the observed morpho-textural changes. Furthermore, the quantified imaging features were not validated against established radiobiological endpoints, and accordingly, their capability to quantitatively evaluate or predict radiation-induced biological responses remains to be confirmed through future studies integrating biological assays. Likewise, the present study did not investigate the independent discriminatory ability of radiomics-derived metrics beyond 3D reconstructed volume and estimated sphericity measurements, which requires to be further explored in future studies with larger datasets and multivariate classification frameworks.

Conclusion

The present study showed that a label-free, morpho-textural radiomics approach allowed non-invasive and longitudinal monitoring of radiation-induced responses in 3D spheroid models using standard brightfield imaging. Through capturing early and progressive morphological variations beyond conventional size-based metrics, the proposed methodology offered a simple yet informative tool for radiobiology studies. As such, this approach holds promise as an imaging-based framework to track evolving morpho-textural responses to irradiation in 3D spheroid models. Nevertheless, a limitation of the present study is that although the extracted radiomic features demonstrated consistency with irradiation-induced growth inhibition trends, direct correlations with specific radiation-induced biological effects were not established. Therefore, the effectiveness and clinical translation potential of the proposed approach require further validation through future studies integrating biological endpoints.

Methods

Cell growth and maintenance

The commercially available Human Embryonic Kidney (HEK-293) cell line (ATCC CRL-1573) was used because it has been extensively used in biomedical research, not only as a robust tool for rapid growth and recombinant protein expression due to its human origin⁵⁹ but also as a model with tumorigenic potential⁶⁰. Moreover, it has been reported that cell survival curves obtained from radiotherapy experiments using 3D spheroids and solid tumors were comparable; thus, spheroids were considered to be in vitro models to mimic basic biological properties of tumor cells in vivo^23,61–63. HEK-293 cells were cultured in DMEM F-12 medium containing 10% Fetal Bovine Serum (FBS), 1% Pen/Strep, 1% non-essential amino acids (NEAA), and 2 mM glutamine, using T75 and T175 flasks in an incubator at with a humidified 5% CO to mimic physiological conditions and ensure consistent cell growth. The cell culture medium was replaced every 48 hours. When the cells reached approximately 80% confluence, they were detached from the flask bottom using 2 mL Trypsin-EDTA (0.25%) sourced from Sigma-Aldrich (St. Louis, MO, USA), resuspended, passaged, and seeded into new culture vessels at a 1:3 ratio to maintain them in an exponential growth phase. The resulting cell suspension was adjusted to cells/mL in complete culture medium and subsequently used for the generation of HEK-293 spheroids through the hanging-drop method. The cell culture preparation protocol for the FaDu spheroids, derived from a subline of the human hypopharyngeal carcinoma cell line FaDu (ATCC HTB-43), was described earlier in detail in the original publication^24,79.

Hanging-drop spheroid formation and experimental imaging workflow

3D HEK-293 spheroids were generated through the hanging-drop method, in which healthy and proliferating cells were transferred and cultured in drops maintained on a superhydrophobic surface, enabling spontaneous, uniform, and physiologically relevant spheroid formation^64,80. This method capitalized on gravity-driven cellular self-assembly within droplets of culture medium, allowing cell-cell interactions mimicking those found in vivo. For hanging-drop formation, a cell suspension (6.0 × 10⁵ cells/mL) was prepared and 30 μL droplets were then dispensed onto the inner surface of an inverted Petri dish lid (10 cm diameter) using a micropipette. Approximately 50 droplets were prepared per lid, with careful adjustment of inter-droplet spacing to prevent droplet contact or merging during incubation. The droplets were retained by surface tension, allowing them to maintain their suspended form. The bottom part of the Petri dish was filled up to half with sterile phosphate-buffered saline (PBS) to maintain humidity and prevent evaporation during incubation. The lid was carefully closed, and the Petri dish was incubated in the inverted position at 37 °C with a humidified 5% incubator. After 48h under standard cell culture conditions, the suspended cells settled at the bottom of each droplet and formed spheroids via intercellular adhesion and natural cell-cell interactions. To ensure reproducibility and repeatability, the complete HEK-293 spheroid production, irradiation, imaging, and radiomics analysis workflow was independently repeated three times (n = 3, independent experimental runs) for each treatment condition (Control, 2 Gy, 4 Gy, and 16 Gy), generating independent spheroid populations in each experimental run. During each experimental run, spheroids were generated on three separate hanging-drop Petri dish lids, each containing approximately 50 droplets. Following spheroid formation, approximately 80–140 spheroids were obtained per treatment condition for downstream experiments. Spheroids demonstrating irregular shapes, droplet merging, or instability were not taken into account for subsequent analyses. The formed spheroids were harvested by gently rinsing them off the Petri dish lid, which was placed at an angle next to the Petri dish, with 3 mL of complete culture medium, minimizing bubble formation and mechanical disruption during transfer. The resulting spheroid suspension was subsequently distributed into 6-well plates, with four macrowells allocated for BF monitoring and two macrowells allocated for FL imaging while maintaining comparable spheroid densities and nutrient availability throughout the experiments. Each BF macrowell contained approximately 15–25 spheroids under comparable culture conditions. Consequently, during each independent experimental run, each treatment condition contributed approximately 60–100 spheroids to the BF imaging set, while 20–50 spheroids per treatment condition were reserved for the FL imaging set. For each treatment condition, empty wells were intentionally left between occupied wells to minimize disturbance and ease imaging and handling procedures.

All HEK-293 spheroids were initially included in longitudinal live BF imaging experiments. To avoid disturbing spheroids maintained for continued longitudinal follow-up until the end of the experiment (D5), spheroids allocated for D2 endpoint Hoechst staining were maintained in separate wells/plates from the main D5 longitudinal cohort. For endpoint Hoechst staining experiments, spheroids were additionally distributed into separate 6-well plates, with two wells occupied per plate while maintaining approximately 15–25 spheroids per well under comparable culture conditions. Hoechst stained HEK-293 spheroids were also monitored by live BF imaging prior to fixation. Particularly, BF images were first acquired from all spheroids at D2, after which selected wells were fixed for Hoechst staining and excluded from subsequent follow-up. The remaining spheroids continued to be monitored under identical experimental conditions until D5, at which point final BF images were acquired prior to fixation and Hoechst staining. Since the experiments terminated at D5, no further follow-up was performed after endpoint staining. This experimental set-up allowed live BF imaging and then Hoechst staining of the same spheroid populations without interference to the live imaging experiments. In order to maintain spheroid stability and prevent partial immobilization during transfer, irradiation, and imaging, 1.5% (w/w) methylcellulose⁶⁴ was added to the culture medium, which increased the viscosity without affecting the viability of spheroids. These conditions prevented spheroids from aggregating or coming into contact with one another that could potentially influence growth, segmentation, reconstructed volume estimations, and texture measurements.

The primary aim of the present study was to characterize irradiation-induced morphological and textural alterations at the spheroid level using image-based radiomics analysis. Therefore, radiomics features were extracted from individual spheroids that met the predefined image-quality and segmentation criteria and were treated as image-analysis units. Across the three independent experimental runs, the final analyzable D0 BF radiomics dataset comprised 83 spheroids in the Control group, 68 spheroids in the 2 Gy group, 87 spheroids in the 4 Gy group, and 65 spheroids in the 16 Gy group. These values represent the number of spheroids included in the dose and time-dependent BF radiomics analysis across all independent experimental runs. Hence, the analyzed spheroids originated from multiple macrowells distributed across independent experimental repetitions rather than from a single well, plate, or hanging-drop preparation.

Given that the objective of this study was image-based radiomics characterization of irradiation-induced spheroid heterogeneity, individual spheroids were retained as image-analysis units for radiomics feature extraction and population-level characterization. The use of three independent experimental runs was intended to verify the reproducibility of irradiation-induced radiomic patterns across independently generated spheroid populations. Although spheroids cultured within the same macrowell were exposed to comparable environmental conditions and may therefore exhibit some degree of within-well similarity, the dataset reflects image-based spheroid-level measurements obtained from multiple wells and independent experimental runs⁸¹. Accordingly, the findings should be interpreted as radiomics-based characterization of spheroid populations reproduced across independent experiments rather than observations derived from a single experimental preparation. Under the present experimental settings, individual spheroids were treated as separate image-based entities for image-based quantitative analysis. Except for irradiation, control spheroids were handled in the same experimental settings as irradiated spheroids to ensure sham-equivalent handling between groups. Representative stages of spheroid formation, immobilization and imaging are shown in Fig. 6.

Figure 6.

Representative stages of spheroid formation by hanging drop method and corresponding imaging/irradiation workflow (Created with BioRender).

Approximately 200 FaDu cells from exponentially growing monolayers were seeded into 96-well agarose-coated plates (200 μL DMEM per well). FaDu spheroids were cultured in liquid overlay as explained earlier²⁴.

Longitudinal radiomics analysis of spheroids

Growth kinetics and structural evolution of the spheroids were monitored for several days by tracking variations in volume, sphericity, entropy, energy and contrast. Longitudinal BF imaging analyses were performed at the spheroid population level within the same treatment wells throughout the experimental period rather than by repeatedly tracking individually identified single spheroids over time. As in the HEK-293 spheroid experiments, BF imaging and subsequent morphometric and texture analyses were performed using an inverted light microscope equipped with a 5 × objective and a high-resolution camera in tile-stitching mode, starting before irradiation and continuing daily throughout the six-day post-irradiation period. To investigate the response of the spheroids exposed to radiotherapy, 3D reconstructed morphological descriptors (volume and sphericity) and entropy were obtained using the open-source AnaSP 3.0 analysis tool, which enables automated extraction of geometric shape features and gray-level textural properties from segmented spheroid masks^34,82. A custom MATLAB pipeline was used to segment the acquired BF images and generate a binary mask assigning a value of 1 to all pixels; these binary masks, together with the corresponding BF images, were provided as input to the analysis workflow. These binary masks, along with their corresponding BF images, were then used as input to the radiomics analysis workflow to compute GLCM-based texture features, including energy and contrast. Following segmentation, each binary mask was automatically inspected and subdivided into its constituent components (spheroid body and any protrusions). The tool then reconstructed each spheroid in 3D through assuming local rotational symmetry around its maximum Feret axis. The enclosed volume was calculated by voxel counting and converting to using the in-plane pixel resolution^34,82.

Sphericity (S) quantifies how closely the shape of the spheroid approximates that of a perfect sphere. In this study, sphericity was estimated in AnaSP according to Eq. (1)^34,83:

1

where Area denotes the pixel-based area of the spheroid and Perimeter gives the length of its boundary in pixels. S is a dimensionless parameter ranging from 0 to 1, in which a value of 1 shows a perfect sphere. Entropy was also computed within the spheroid mask as a textural descriptor to capture intra-spheroid intensity heterogeneity, quantifying gray-level variation and spatial disorder as a complementary metric to volume and sphericity^37,38.

GLCM analysis represents a second-order statistical texture method capturing spatial correlations between neighboring pixel intensities in grayscale images, such as BF images. A GLCM is defined by counting how often pairs of gray levels p(i, j) occur at a given displacement distance (d) and orientation (θ ), forming a matrix of size N g × N g, in which Ng refers to the number of quantized gray levels. In this study, the original BF images were 16-bit grayscale images. Thus, prior to GLCM computation, the images were quantized into gray levels to optimize preservation of morpho-textural information, computational efficiency, and noise sensitivity^39,84. Since GLCM-derived features are sensitive to gray-level discretization, a fixed quantization level was maintained throughout all analyses to ensure reproducibility and comparability of the quantified features. GLCMs were calculated using a pixel offset distance of d = 1 pixel in a set of four orientations (, , , and ). The resulting GLCM feature values were obtained by averaging the measurements across all four directions, consistent with standard Haralick-based texture analysis approaches, due to spatial directional dependency of textural patterns^39,84. The original GLCM count matrices were normalized to probability matrices during feature calculation, such that the sum of all matrix elements was equal to 1, leading to normalized co-occurrence probabilities p(i,j). Accordingly, three GLCM-based metrics, entropy, energy and contrast were evaluated in this study to quantify radiation-induced textural alterations in spheroids.

Entropy measures the degree of randomness in image texture by reflecting variations in intensity distribution. Entropy increases when all probabilities p(i, j) are uniformly distributed, and decreases as the values of p(i, j) become more heterogeneous^37–39. Accordingly, a homogeneous spheroid will result in a lower entropy value, whereas an inhomogeneous spheroid will lead to a greater entropy value.

2

Energy quantifies the degree of image homogeneity. A small number of relatively high p(i, j) values refers to a homogeneous image including only a few gray levels. Thereby, the sum of squares will be high. Higher energy values correspond to smoother gray-level transitions and greater textural uniformity within the spheroids^39,85.

3

Contrast measures local intensity variations within an image emphasizing that this measure favors contributions from p(i, j) away from the diagonal. It is strongly associated with the difference between neighboring pixel values, particularly at small displacement distances. Images showing pronounced intensity fluctuations thus yield higher contrast values^39,86.

4

The performed GLCM feature calculations followed standard Haralick-based texture definitions and standardized GLCM processing procedures, including gray-level quantization, directional averaging, and matrix normalization. While the custom-developed MATLAB-based GLCM analysis was not directly cross-validated with standardized radiomics platforms such as PyRadiomics⁸⁷ and Imaging Biomarker Explorer (IBEX)⁸⁸ in the present study, the feature extraction pipeline was configured in accordance with commonly used radiomics methodologies.

In addition, treatment response was evaluated by computing growth inhibition. Fold-change in volume was first defined as the relative increase in spheroid 3D reconstructed volume between the baseline time point and the end of the monitoring period. This fold-change metric was used to normalize growth kinetics across spheroids with varying initial sizes and to allow direct comparison between non-irradiated and irradiated groups. This normalization step minimized variability caused by biological differences between spheroids, ensuring more reliable comparisons. Then, growth inhibition (GI) was quantified using a baseline and control normalized formulation. Particularly, spheroid volumes at time point t were first normalized to their corresponding baseline spheroid volumes at , and the normalized growth of irradiated spheroids was then compared to that of the control group. GI was figured out to evaluate the radiotherapy effect at the last time point (D5) with complete observations of both control and treatment groups, as given in Eq. (5)^89,90:

5

where and denote the median spheroid volumes of the irradiated and non-irradiated (control) groups at time point t, respectively, while and represent the corresponding median baseline volumes measured at the initial time point .

Spheroids were irradiated on the third day of formation that was defined as D0. All spheroids within each cell population were monitored every 24 h under four conditions: Control (no irradiation), 2 Gy, 4 Gy, and 16 Gy. To assess growth kinetics, the change in spheroid volume, sphericity and entropy from the initiation of the experiment D0 to the end of the monitoring period Day 5 (D5, referring to 6th day after irradiation) was used to obtain a growth dynamics. Moreover, all spheroid volume data were expressed as relative volume, normalized to the group-specific median volumes of spheroids at D0 to decrease the impact of outlier values and offer a more proper reference to make a comparison across groups. The relative volume was computed using a baseline-normalized formulation:

6

where refers to the first available imaging time point (D0 for HEK-293 spheroids and D5 for FaDu spheroids).

To biologically validate the radiomics results derived from BF imaging, spheroids were fixed, and all nuclei were stained by Hoechst (Sigma-Aldrich St. Louise, MO, USA) before fluorescence image acquisition . Spheroids were incubated for 30 minutes. To further assess the robustness and generalizability of the analysis workflow, an open-source BF image dataset of X-ray irradiated FaDu tumor spheroids was also investigated. This dataset includes spheroids derived from a subline of the human hypopharyngeal carcinoma cell line FaDu (ATCC HTB-43)^24,91, imaged at multiple time points (D0, D3, D5, and D7) following single-fraction X-ray exposures of 5 Gy, 7.5 Gy, or 20 Gy⁷⁹, together with non-irradiated FaDu spheroids. For these spheroids, volume, sphericity, and entropy were quantified using the same pipeline to directly compare radiation-induced growth and structural response patterns between HEK-293 and FaDu spheroids across radiation dose and follow-up time points.

Spheroid segmentation and radiomics analysis

HEK-293 spheroids transferred into 6-well plates were imaged at multiple positions using a 5 × objective with tile-scanning to capture the entire cell population within each well. All BF images used for subsequent radiomics analysis were acquired using the same optical configuration and magnification settings to maintain a spatial resolution of approximately throughout the study. BF images were pre-processed by applying pseudo flat-field correction to compensate for radial fall-off attenuation in image intensity⁹², followed by image inversion to enhance contrast. For spheroid segmentation, a circular region of interest was manually drawn to guide the active contour algorithm, enabling semi-automated boundary detection. A custom MATLAB script (R2023b, MathWorks, Natick, MA, USA) was then used to generate the edge-detected spheroid outlines and produce the corresponding binary mask images that served as the basis for subsequent quantitative analysis without additional boundary erosion. The generated binary masks were organized into datasets for use in the AnaSP 3.0 software to enable quantitative estimation of reconstructed 3D volume, sphericity, and entropy estimations. The same image processing workflow was applied to FaDu spheroids, with all BF images recorded under the same optical settings corresponding to a spatial resolution of approximately . In addition, no predefined minimum spheroid size or mask area threshold was applied for inclusion in either spheroid model. However, spheroids showing unstable segmentation or insufficient structural integrity were excluded from the analysis pipeline. A representative workflow of the spheroid image processing pipeline is shown in Fig. 7.

Figure 7.

Steps of spheroid image processing in MATLAB. Brightfield images of HEK-293 spheroids were first pre-processed using pseudo flat-field correction and contrast inversion, followed by semi-automated segmentation with active contour. The resulting edge-detected spheroid masks by active contour were then used in AnaSP 3.0 software to extract morphological and texture parameters³⁴.

Irradiation protocol

HEK-293 spheroids kept in 6-well plates were irradiated by a total and single-dose application of 2 Gy (conventional), 4 Gy (hypofractionated), and 16 Gy (ablative) in a single fraction by a Linear Accelerator (LINAC) device (Elekta Versa HD, Elekta, Crawley, UK). Following a protocol similar to that used in our previous study⁹³, this study used a 6 MV photon beam with a flattening filter to deliver doses of 2 Gy, 4 Gy and 16 Gy to the spheroids. Spheroid cultures were located at the isocenter with robust buildup material to confirm dose delivery at and the field size was adjusted to fully encompass the well region of the culture plates. Single-fraction doses of 2 Gy, 4 Gy, and 16 Gy were delivered in one exposure each, whereas non-irradiated control spheroids were kept under the same environmental conditions. During irradiation, the 6-well plates were submerged in a custom-made phantom to ensure homogeneous distribution of radiation and simulate body structure. The phantom was scanned using Computer Tomography (CT) with a slice thickness of 2 mm before treatment. The plate keeping the spheroids was considered as the Gross Tumor Volume (GTV) for dose-definition purposes. In addition to conventional fractionation doses such as 2 Gy and 4 Gy, a single fraction ablative dose of 16 Gy, typically used in Stereotactic Radiosurgery/Radiotherapy (SRS/SRT) applications, was also applied. All irradiations were performed using an external RT unit (Versa HD linear accelerator, Elekta, Crawley, UK). The Monaco Treatment Planning System took over the dose distributions of the current study. Through recommended doses using fields of radiation Anterior-Posterior/Posterior-Anterior (AP/PA), the volume of the GTV complied with 95 / 100(95 %). Using the collapsed cone algorithm Treatment Planning System (TPS, Monaco 5.11.02v), the monitor unit (MU) was calculated. The MU number was determined using fields using the collapsed cone calculation technique to account for inhomogeneities to deliver the radiation doses of , and 16 Gy to the HEK-293 spheroids in the macrowells. LINAC output was corrected to confirm differences in doses at smaller than 2% differences following TRS 398⁹⁴ before the irradiation. Using a 0.6 cc farmer-type ionization chamber, delivered doses were verified for each experiment. The phantom positioning was confirmed to ensure spatial correspondence between the wells and the region delineated in the CT-based irradiation plan. The plate positioning within the phantom was checked through matching the Digitally Reconstructed Radiograph (DRR) with the corresponding kV images.

For the open-source FaDu dataset included in this study, spheroids were irradiated in 96-well plates using single X-ray doses in the 0-25 Gy range (200 kV; 0.5-mm Cu filter; Yxlon Y.TU 320; Yxlon International, Hamburg, Germany) as reported in the original study^24,79. From this dataset, Control, 5 Gy, 7.5 Gy, and 20 Gy dose exposures were selected for further analysis.

Statistical analysis

Each experimental condition for HEK-293 spheroids was performed in triplicate to minimize potential cell culture related variability and ensure reproducible morphological and textural measurements. All morphological and entropy-derived measurements were given using the median and interquartile range, consistent with the non-parametric nature of the dataset. Statistical analyses were performed in MATLAB (R2023b, MathWorks, Natick, MA, USA). Group differences between controls and irradiated spheroids were investigated using the Kruskal-Wallis test, followed through Dunn’s Šidák-corrected multiple comparisons where appropriate. Effect sizes (Cliff’s δ) were calculated to quantify the magnitude and direction of differences between groups. Since GI produced a single summary value per treatment group, GI% were directly reported instead of performing statistical comparisons. The same non-parametric methodology was used to look at dose-dependent variations among Control, 5 Gy, 7.5 Gy and 20 Gy groups. P<0.05 was set as a significance level. For all experiments, statistical significance was defined as follows: *P<0.05; **P<0.01; ***P<0.001.

Author contributions

I.D and K.A conceived, designed, and performed the experiments, analyzed the results, and wrote the manuscript. All authors reviewed the manuscript.

Funding

This study received no specific funding.

Data availability

The dataset measured and analyzed during this study is available from the corresponding author upon reasonable request.

Declarations

Competing interests

The authors declare no competing interests.