Predicting neoadjuvant chemotherapy response in osteosarcoma using T2-based MRI delta-radiomics: a retrospective study

Abstract

Objective

This study aimed to investigate the application of T2-based MRI delta-radiomics as a novel predictive tool for neoadjuvant chemotherapy (NACT) response in patients with osteosarcoma.

Materials and methods

We retrospectively analyzed data from 152 patients with pathologically confirmed osteosarcoma who underwent NACT at our institution. Axial T2-weighted MRI sequences were acquired both at baseline (pre-NACT) and after NACT (post-NACT). After image segmentation and preprocessing, 1158 radiomic features were extracted from the T2-weighted images. We developed and compared four models: the conventional quantitative imaging features-based model (CQIF model), the pre-NACT radiomics model, the post-NACT radiomics model, and the Delta-Radiomics model. Model performance was assessed using the area under the receiver operating characteristic curve (AUC) and accuracy (ACC).

Results

Based on histopathological assessment, patients were divided into two groups: good responders (n = 57) and poor responders (n = 95). Significant differences in change rates for tumor diameter and volume were observed between the two groups (P < 0.001). The Delta-Radiomics model demonstrated superior predictive performance compared to other models, achieving an AUC of 0.796, ACC of 0.756, sensitivity of 0.529, specificity of 0.893, PPV of 0.750, and NPV of 0.758 in the test set. However, the Delong test revealed no significant differences among these models, except between the Post-NACT and Delta-Radiomics models (P < 0.05).

Conclusions

T2-based MRI delta-radiomics showed strong predictive value for NACT response in patients with osteosarcoma. This model holds potential for guiding clinical decision-making and improving patient management by identifying responders early in the treatment course.

Introduction

Osteosarcoma is the most common primary bone malignancy, predominantly affecting adolescents and young adults [1−4]. Characterized by aggressive growth and a high propensity for metastasis, osteosarcoma poses significant treatment challenges, including a notable rate of resistance and recurrence despite intensive chemotherapy regimens [5]. The standard treatment for osteosarcoma generally involves a combination of surgical resection, neoadjuvant chemotherapy (NACT), and adjuvant chemotherapy. The current first-line chemotherapy regimen typically consists of a combination of methotrexate, doxorubicin, and cisplatin [6]; however, patient outcomes can vary considerably [7]. Therefore, accurately predicting treatment response is crucial to optimize individual treatment plans and improve patient outcomes.

Magnetic resonance imaging (MRI) offers superior soft tissue contrast and can capture dynamic changes in tumor characteristics over time, making it one of the most widely used assessment methods in clinical practice. According to the Response Evaluation Criteria in Solid Tumors, version 1.1 (RECIST 1.1) criteria [8], traditional imaging assessments typically measure the change in tumor diameter before and after chemotherapy. However, changes in tumor diameter alone have limited value in accurately assessing treatment efficacy, particularly in osteosarcoma, where tumor size may not always correlate with underlying biological changes [9]. Radiomics, the extraction of quantitative features from medical images, has emerged as a promising approach to deepen our understanding of tumor biology [10−12]. Among these techniques, delta-radiomics evaluates changes in radiomic features over time and shows significant potential in predicting treatment outcomes [9, 13−15]. By quantifying differences in tumor characteristics from MRI scans taken before and after chemotherapy, delta-radiomics can identify patterns associated with treatment response. Specifically, T2-weighted MRI delta-radiomics has demonstrated its utility in predicting pathological complete response in high-grade soft tissue sarcoma (STS) patients undergoing neoadjuvant therapy [13−15]. Despite these advancements, many previous studies have focused solely on pre-treatment MRI features [16−18], often neglecting the critical comparison of baseline and post-chemotherapy scans. This highlights the need for further investigation into the changes captured through delta-radiomics to enhance predictive accuracy.

The aim of this study was to investigate the potential of T2-based MRI delta-radiomics as a predictive tool for evaluating NACT response in osteosarcoma patients. By analyzing changes in tumor characteristics derived from pre- and post-NACT MRI scans, we sought to develop a robust model that could improve treatment prediction and personalize therapy for osteosarcoma patients. This approach is expected to not only enhance early response assessment but also provide valuable insights into the dynamic nature of tumor response to chemotherapy, ultimately improving clinical decision-making and patient outcomes.

Materials and methods

Participants and data acquisition

This study was approved by the local ethics committee of our hospital, and informed consent was waived due to the retrospective nature of the study. A total of 152 patients with pathologically proven osteosarcoma who underwent NACT at our institution from August 2013 to October 2023 were retrospectively analyzed. The inclusion criteria were as follows: (1) newly diagnosed osteosarcoma patients, with no prior treatment; (2) availability of axial T2-weighted MRI sequences taken at baseline (pre-NACT) and after NACT (post-NACT); (3) confirmed diagnosis of osteosarcoma based on pathology reports. Exclusion criteria included patients without axial T2-weighted MRI images, poor image quality, or lack of pathological confirmation regarding the efficacy of NACT.

All patients received 3 to 6 cycles of NACT prior to surgery, typically involving doxorubicin, cisplatin, methotrexate, and ifosfamide [18, 19]. The efficacy of NACT was assessed primarily through the tumor necrosis rate (TNR) following treatment. Patients with a TNR of 90% or higher were classified as good responders, while those with lower rates were designated as poor responders [20].

Additionally, clinical data—including sex, age, and tumor location—along with conventional quantitative imaging features (tumor diameter, tumor volume, and relative signal intensity) were analyzed. Figure 1 illustrates the study flowchart.

Fig. 1.

The workflow of this study

Image acquisition and data preprocessing

All MRI images were obtained on a 1.5T or 3.0T scanner (1.5T uMR 588, United Imaging; Philips Ingenia 3.0T; SIGNA EXCITE 1.5T, GE Healthcare; 3.0T Discovery 750/750 W, GE Healthcare) in our hospital following various protocols depending on the sites of lesions. Each patient included axial T2-w, and the parameters were as follows: TR/TE = 2060 ~ 5637 / 63.4 ~ 120.6 ms, slice thickness = 5.0 ~ 10.0 mm, matrix = 320 × 179 ~ 400 × 368, field of view = 14 × 14 cm ~ 44 × 44 cm.

The preprocessing of the datasets involved several key steps: First, all datasets were isotropically resampled in three dimensions using linear interpolation, with x-, y-, and z-spacing adjusted to 1 mm × 1 mm × 1 mm. Next, a Gaussian filter with a standard deviation of 0.5 was applied for signal smoothing. To address inhomogeneity artifacts and ensure consistent intensity distribution, bias field correction and intensity standardization (gray level discretization from 0 to 255) were also performed [21].

Tumor segmentation and conventional quantitative imaging features analysis

Three-dimensional regions of interest (ROIs) for each patient were defined on T2-weighted images using a response assessment platform [22, 23] based on WEASIS software (https://weasis.org/en/index.html). This delineation was performed by a musculoskeletal radiologist with six years of experience and subsequently validated by a senior musculoskeletal radiologist with fifteen years of experience. Each layer focuses on the soft tissue mass component of the lesion, intentionally excluding any evident areas of edema.

Three conventional quantitative imaging features were calculated: the tumor’s diameter, tumor volume, and relative signal intensity. First, the maximum diameter and volume of the tumor were automatically computed upon completion of the delineation process and pressing the confirmation button. Relative changes in tumor maximum diameter and volume were then calculated by subtracting the post-NACT values from the pre-NACT values and dividing the results by the pre-NACT values.

Additionally, we delineated a ROI in the muscle adjacent to the lesion during the pre-NACT phase and copied this ROI to the corresponding slice of the same muscle post-NACT. To obtain the relative signal intensity, we divided the signal intensity of the lesion by the signal intensity of the adjacent muscle. The relative change in signal intensity was similarly calculated by subtracting the pre-NACT value from the post-NACT value and dividing by the pre-NACT value.

Radiomics feature extraction and preprocessing

In total, 1158 well-defined MRI radiomic features, encompassing both 2D and 3D characteristics, were computed for each segmented osteosarcoma mask using in-house feature extraction package, the Columbia Image Feature Extractor [24], which was implemented on the MATLAB 2016b platform (MathWorks, Natick, MA). These 1158 extracted features were encompassed shape descriptors (e.g., eccentricity and compactness), size metrics (e.g., volume and largest diameter), intensity contrast parameters (e.g., Laplacian of Gaussian filtering), boundary sharpness attributes (e.g., sigmoid slope), and texture patterns (e.g., Gray-Level Cooccurrence Matrix) [25]. Delta-radiomic features were calculated by subtracting the features derived from the pre-NACT value from those obtained post-NACT, and then dividing the result by the pre-NACT value.

Normalization of the extracted features is the first step before performing feature selection. Outliers were replaced with the median of the corresponding variance vector once their values exceeded three times the range of the mean and standard deviation. Additionally, the data was standardized within a specific interval. The standardized formula is as follows: (fi − u)/ σ, where fi represents a single data point, u is the mean of the data column, and σ is the standard deviation of the data column.

Radiomics feature selection

Univariate Wilcoxon tests, correlation analysis, and multivariate stepwise regression were employed to minimize feature redundancy and selection bias, with highly correlated features being excluded. Radiomics score (Radscore) was calculated for each patient via a linear combination of selected features that were weighted by their respective coefficients.

Model building

Finally, the most significant features were investigated to construct the radiomics model on the basis of logistic regression. 152 patients were randomly divided into a training set (n = 107) and a test set (n = 45) at a ratio of 7:3. We developed and compared four different models using a variety of features, including the conventional quantitative imaging features-based model (CQIF model), the pre-NACT radiomics model (Pre-NACT model), the post-NACT radiomics model (Post-NACT model), and the delta-radiomic features-based model (Delta-Radiomics model). Models were trained using the repeated 3-fold cross-validation method in the training set, and estimation performance was evaluated in the validation set.

Validation of the radiomics models

The performance of different models was assessed using the area under the receiver operating characteristic curve (AUC) and accuracy (ACC). The calibration plots were used to graphically investigate the performance characteristics of different nomograms. Decision curve analysis (DCA) was used to assess the clinical usefulness of the models. Figure 1 illustrates the workflow of this study.

Statistical analyses

Statistical analyses were performed with R software (R Core Team, Vienna, Austria) version 3.4.3. The t-test or Mann–Whitney U test was performed to compare continuous variables, while chi-squared or Fisher’s exact test was used for classify variables between groups. All statistical tests were two-sided, a Bonferroni-corrected P value was used to identify the feature significance of multiple comparisons.

Results

Participant characteristics

A total of 152 patients diagnosed with osteosarcoma were included in this study. The demographic and clinical characteristics of the participants are summarized in Table 1. The median age at diagnosis was 14.0 years (interquartile range: 12.0 to 18.0), with a gender distribution of 63 females and 89 males. Following histopathological assessment post-surgery, patients were classified into two groups: the good response group (n = 57) and the poor response group (n = 95), based on the percentage of necrosis observed. No significant differences were found in age, sex, or tumor location between the two response groups (P > 0.05).

Table 1.

Clinical characteristics of patients

Variable	Poor response (n = 95)	Good response (n = 57)	χ2 / Z value	P value
Sex			χ2  = 1.520 a	0.218
Female	43 (45.3%)	20 (35.1%)
Male	52 (54.7%)	37 (64.9%)
Age (years)	14.0 (12.0, 21.0)	14.0 (11.0, 16.5)	Z = -1.046 b	0.295
Tumor location			χ2  = 2.671 c	0.449
Femur	51 (53.7%)	34 (59.6%)
Tibiofibula	26 (27.4%)	18 (31.6%)
Humerus	8 (8.4%)	2 (3.5%)
Pelvis and sacrum	10 (10.5%)	3 (5.3%)

Note: a, chi-squared test. b, Mann–Whitney U test. c, Fisher’s exact test

Conventional quantitative imaging features analysis

Pre-NACT MRI scans showed a median tumor diameter of 65.73 mm, with an interquartile range of 52.24 mm to 79.95 mm. The overall range of tumor diameters was wide, spanning from 20.13 mm to 146.86 mm. Post-NACT MRI scans showed a significant reduction in tumor size, with a median diameter of 60.71 mm (interquartile range: 48.27 mm to 78.22 mm).

Pre-NACT MRI scans indicated a median tumor volume of 117840.00 (49707.35, 203825.25) mm³, with a range from 2887.50 mm³ to 1270360.00 mm³. Post-NACT MRI scans showed a significant reduction in tumor volume, with a median volume of 99218.30 (40136.43, 202925.00) mm³.

We found no significant statistical differences in tumor maximum diameter, volume, and relative signal intensity between the good response and poor response groups in the pre-NACT phase (P > 0.05). Similarly, there were no significant differences in tumor maximum diameter and volume between the two groups in the post-NACT phase (P > 0.05). However, relative signal intensity showed a statistically significant difference (Z = − 2.110, P = 0.035). In terms of change rates, there were statistically significant differences in the tumor maximum diameter and volume between the two groups (Z = − 4.653, Z = − 5.003, P < 0.001), while the change rate of relative signal intensity did not show a significant difference (P > 0.05) (Table 2).

Table 2.

Conventional quantitative imaging features of good- and poor-response groups before and after chemotherapy

	Pre-NACT				Post-NACT					(Post-NACT - Pre-NACT) / Pre-NACT
	Good response	Poor response	Z a	P	Good response	Poor response	Z a	P		Good response	Poor response	Z a	P
Tumor diameter (mm)	66.51 (54.69, 80.73)	64.35 (47.48, 79.85)	−1.418	0.156	58.26 (47.80, 74.27)	63.39 (48.29, 83.23)	−1.454	0.146		−0.12 (− 0.24, − 0.05)	0.01 (− 0.12, 0.20)	−4.653	0.000
Tumor volume (mm3)	116621.00 (63719.20, 252808.50)	119103.00 (35162.50, 176258.00)	−1.437	0.151	73664.40 (41069.35, 154893.00)	118899.00 (39729.50, 235594.00)	−1.452	0.147		−0.33 (− 0.51, − 0.06)	0.17 (− 0.31, 0.68)	−5.003	0.000
Relative signal intensity	1.88 (1.46, 2.56)	2.18 (1.79, 3.30)	−1.730	0.084	1.58 (1.11, 2.63)	1.95 (1.53, 2.71)	−2.110	0.035		−0.13 (− 0.36, 0.18)	−0.15 (− 0.36, 0.20)	−0.093	0.926

Note: a, Mann–Whitney U test

For the CQIF model, only the feature diameter_change was retained for model building after feature selection.

Radiomic feature analysis

For the pre-NACT MRI scans, 3 features, including Shape_SI5_Para_2, Sigmoid_Offset_Std_Para_2 and Spatial_Correlation_Para_7, were selected to build the Pre-NACT model.

For the post-NACT MRI scans, only one feature (GLCM_IMC1_Para_7) was selected to build the Post-NACT model.

Five features—Sigmoid_Offset_Std_Para_5, DWT1_LL_Para_1, GTDM_Busyness_Para_5, Laws_8_Z_Boundary_Para_2, and GLCM_IMC2_Para_8—were selected to build the Delta-Radiomics model. Sigmoid_Offset_Std_Para_5, a morphological/statistical feature, reflects changes in the tumor’s shape or boundary, specifically the offset and standard deviation of a Sigmoid function, which characterizes the tumor’s contrast and complexity based on gray-level variations in the image. DWT1_LL_Para_1 is derived from the Discrete Wavelet Transform (DWT), focusing on the low-frequency sub-band (LL), capturing local features at multiple scales to reveal details about the tumor’s structure. GTDM_Busyness_Para_5 is computed from the Gray Level Distance Matrix (GTDM), which measures the “busyness” or complexity of the tumor’s texture, with higher values indicating greater heterogeneity and biological activity. Laws_8_Z_Boundary_Para_2, based on Laws’ texture energy method, analyzes texture at different scales, specifically reflecting boundary characteristics and helping to identify irregular tumor borders, which can be useful for differentiating tumor types. Finally, GLCM_IMC2_Para_8, a feature from the Gray-Level Co-occurrence Matrix (GLCM), quantifies the contrast and uniformity of the tumor texture, providing valuable information on tumor heterogeneity and its spatial relationships. Table 3 shows the features included in the Delta-Radiomics model after multivariate stepwise regression. Figure 2 the violin plot of the distribution of retained feature values for Delta-Radiomics model.

Table 3.

The results of the multivariable logistic regression analyses for the Delta-Radiomic model

Variable	Coefficient	OR (95% CI)	P
Intercept	-1.047	-	0.000
Sigmoid_Offset_Std_Para_5	-0.844	0.430(0.179–0.863)	0.039
DWT1_LL_Para_1	0.629	1.875(1.125–3.363)	0.022
GTDM_Busyness_Para_5	-1.314	0.269(0.089–0.640)	0.009
Laws_8_Z_Boundary_Para_2	2.755	15.724(2.387-123.876)	0.006
GLCM_IMC2_Para_8	-0.892	0.410(0.206–0.725)	0.005

Note: OR, odds ratio; CI, confidence interval

Fig. 2.

Violin plot of retained radiomics feature values for Delta-Radiomics model. In panels A-E, the shape of each violin plot illustrates the density of data at various feature values, emphasizing variations among the retained parameters. The x-axis represents the poor response (no) and good response (yes) groups, while the y-axis shows the medians and interquartile ranges of the various features

Performance of radiomics models

The Delta-Radiomics model demonstrated superior predictive performance compared to other models, achieving an AUC of 0.813, ACC of 0.841, sensitivity of 0.675, specificity of 0.940, PPV of 0.871, and NPV of 0.829 in the training set. In the test set, the model achieved an AUC of 0.796, ACC of 0.756, sensitivity of 0.529, specificity of 0.893, PPV of 0.750, and NPV of 0.758 (Fig. 3; Table 4).

Fig. 3.

ROC curves of different models in the training set (A) and test set (B). The Delta-Radiomics model (red line) demonstrated superior predictive performance compared to the other models

Table 4.

Performance of different models in training set and test set

	AUC	ACC	Sensitivity	Specificity	PPV	NPV
CQIF model	0.728(0.729)	0.682(0.711)	0.850(0.706)	0.582(0.714)	0.548(0.600)	0.867(0.800)
Pre-NACT model	0.704(0.697)	0.617(0.556)	0.875(0.824)	0.463(0.393)	0.493(0.452)	0.861(0.786)
Post-NACT model Delta-Radiomics model	0.631(0.519) 0.813(0.796)	0.570(0.422) 0.841(0.756)	0.775(0.706) 0.675(0.529)	0.448(0.250) 0.940(0.893)	0.456(0.364) 0.871(0.750)	0.769(0.583) 0.829(0.758)

Note: CQIF model, conventional quantitative imaging features-based model; Pre-NACT model, pre-NACT radiomics model; Post-NACT model, post-NACT radiomics model; AUC, area under curve; ACC, accuracy; PPV, positive predictive value; NPV, negative predictive value. Training set, in front of the brackets. Test set, in brackets

The CQIF model achieved an AUC of 0.728, an ACC of 0.682, a sensitivity of 0.850, a specificity of 0.582, a PPV of 0.548, and a NPV of 0.867 in the training set. In the test set, it had an AUC of 0.729, an ACC of 0.711, a sensitivity of 0.706, a specificity of 0.714, a PPV of 0.600, and a NPV of 0.800. These results indicate that tumor diameter is a valuable predictor of treatment response.

The Pre-NACT model achieved an AUC of 0.704 and an ACC of 0.617 in the training set, with sensitivity at 87.5% and specificity at 46.3%. In the test set, it recorded an AUC of 0.697, an accuracy of 0.556, sensitivity of 82.4%, and specificity of 39.3%.

The Post-NACT model achieved an AUC of 0.631 and an ACC of 0.570 in the training set, with sensitivity at 77.5% and specificity at 44.8%. In the test set, it showed an AUC of 0.519, ACC of 0.422, sensitivity of 70.6%, and specificity of 25.0%.

However, the Delong test revealed no significant differences among these models, except between the Post-NACT and Delta-Radiomics models (P < 0.05). Figure 4 shows the calibration curves and DCA curves of the different models in the test set.

Fig. 4.

The calibration curve and DCA curve of different models in the test set. Panel A shows that all models had good predicted probabilities that accurately reflected the observed outcomes, except for the Post-NACT model (blue line). Panel B shows the DCA curve, where the Delta-Radiomics model (red line) demonstrated the best performance, yielding significant clinical net benefits

Discussion

Osteosarcoma remains a challenging malignancy, particularly in adolescents and young adults, where effective treatment strategies are essential for improving outcomes. In this study, we explored the use of T2-based MRI delta-radiomics as a novel predictive tool for assessing NACT response in osteosarcoma patients. Our findings demonstrate that delta-radiomic features provide significant predictive value, outperforming traditional models based solely on tumor diameter.

The retrospective analysis of 152 patients revealed no significant demographic differences between those classified as good responders and poor responders based on histopathological assessment. This suggests that factors such as age, sex, and tumor location may not be sufficient predictors of NACT response, consistent with previous literatures [9, 12, 26].

In this study, we observed significant changes in tumor maximum diameter and volume between the response groups, underscoring the critical importance of monitoring these morphological parameters during treatment [27]. Interestingly, while relative signal intensity showed a significant difference between the groups in the post-NACT phase, its change rate did not yield significant results. This discrepancy suggests that, although relative signal intensity on T2-weighted images can be influenced by treatment, it may not reliably indicate treatment response when assessed based on change rates. This further implies that specific morphological changes may provide a more accurate reflection of treatment response than variations in signal intensity alone. Kanthawang et al. [28] found that tumor volume and maximum axial diameter can serve as independent predictors of chemoresistance in osteosarcoma based on baseline MRI findings. In contrast, Yildirim et al. [29] identified the largest enhancing component on contrast-enhanced T1-weighted images—not tumor volume or intramedullary component size—as a promising preoperative indicator of NACT response in osteosarcoma. Future studies should explore the underlying biological mechanisms behind these observations, as this could enhance our understanding of tumor behavior in response to chemotherapy and inform the development of more effective treatment strategies.

Radiomic features offer a more detailed analysis of tumors compared to conventional imaging, capturing complex and often invisible characteristics. By including features from various domains—such as morphological changes (Sigmoid_Offset_Std), wavelet features (DWT1_LL), texture analysis (GTDM_Busyness, GLCM_IMC2), and boundary irregularity (Laws_8_Z_Boundary)—radiomics provides a comprehensive view of tumor heterogeneity [25]. This multi-dimensional approach allows for the capture of both macroscopic changes (e.g., size and volume) and microscopic variations (e.g., texture and boundary features), making it a powerful tool for predicting treatment response. These features can reflect underlying biological changes, such as alterations in cellular architecture and metabolic activity following chemotherapy. The strong correlation between specific delta-radiomic features and histopathological response further supports the idea that MRI-derived metrics can serve as surrogate markers for tumor biology. The inclusion of delta-radiomic features, which reflect dynamic changes in tumor characteristics over the course of treatment, likely accounts for the superior performance of the Delta-Radiomics model. These features capture subtle patterns in tumor heterogeneity that conventional imaging may miss, providing deeper insights into the biological processes driving treatment response [30].

Importantly, the Delta-Radiomic model showed superior predictive performance when compared to other models, highlighting its ability to capture subtle tumor changes that conventional assessments may miss [9]. By analyzing a high-dimensional array of features from T2-weighted MRI images, this model provides earlier and more accurate predictions of treatment response, crucial for tailoring personalized patient management. Previous studies have suggested that the T2-weighted MRI delta-radiomics approach can enhance early response assessment in STS patients [13 − 15]. Zhang et al. [12] demonstrated that radiomics models based on pre-NACT, post-NACT, and combined MRI could effectively predict chemotherapy response in osteosarcoma patients. Additionally, Lin et al. [9] showed that delta-radiomics models based on computed tomography could be utilized for individualized pathological response evaluation following preoperative chemotherapy in osteosarcoma patients. In this study, we are the first to identify T2-based MRI delta-radiomics as a novel predictive tool for assessing NACT response in osteosarcoma patients. The clinical significance of our findings lies in the potential to identify non-responders early, allowing for timely adjustments to treatment. This could mitigate the risks associated with ineffective therapies and minimize chemotherapy-related side effects. By incorporating delta-radiomics into clinical practice, oncologists could personalize treatment strategies and improve patient outcomes [31]. Furthermore, while the delta-radiomics model demonstrated overall strong discriminatory power, its moderate sensitivity for identifying non-responders remains a key limitation for immediate clinical translation. The precise identification of this subgroup is crucial for early therapeutic adjustment. The current model prioritizes a higher specificity to minimize false positives, which is also clinically vital to avoid incorrectly altering effective therapy. Therefore, its utility as a standalone tool for reliably triaging all non-responders is limited. Our findings should be interpreted as a promising proof-of-concept, and future efforts are needed to improve sensitivity, potentially through advanced feature selection or multi-modal integration, before clinical application can be realized. Additionally, DCA showed that the Delta-Radiomics model provided the highest clinical net benefit, suggesting its potential to guide treatment decisions and optimize clinical outcomes by accurately identifying patients most likely to benefit from therapy. The Delong test confirmed that the Post-NACT model performed significantly worse than the Delta-Radiomics model, emphasizing the added value of dynamic, delta-radiomic features derived from pre- and post-treatment scans. Although the Pre-NACT and Post-NACT models showed more modest performance, particularly in terms of specificity and accuracy, they highlight the importance of incorporating tumor feature changes over time for more accurate predictions of treatment response [13].

Our study has several limitations. First, the retrospective, single-center design and relatively modest sample size may limit the statistical robustness of our findings and increase the risk of random error or selection bias. Although rigorous internal validation was employed, the generalizability of the proposed delta-radiomics model remains constrained. Therefore, our results should be interpreted with caution, and further validation in larger, prospective, multi-center cohorts is essential to confirm the model’s reproducibility and clinical applicability across diverse populations and imaging protocols. A further limitation pertains to the potential for overfitting and model optimism, introduced by the high-dimensional radiomic feature space relative to our cohort size. Despite employing a staged feature reduction strategy, the entire selection process was conducted within a single, internal train-test split. Consequently, the generalizability and stability of the final feature set remain uncertain. The performance metrics reported here should therefore be interpreted as preliminary and derived under internal validation conditions; robust external validation on fully independent, multi-center cohorts is required to confirm the model’s true predictive utility and to mitigate the risk of overfitting. Third, the retrospective inclusion of MRI data acquired from multiple vendors and field strengths with heterogeneous acquisition parameters represents another key limitation. Although standardized preprocessing and feature normalization were applied to reduce scanner-related effects, these methods cannot fully eliminate inherent technical variability. This residual heterogeneity may affect feature stability and model generalizability, highlighting the need for prospective harmonization in future studies. Furthermore, the present model is based solely on radiomic features. The exclusion of clinical and molecular variables, which may hold complementary predictive value, could limit the model’s comprehensiveness and immediate clinical applicability. The integration of such multimodal information represents an important direction for future research. Longitudinal studies that examine radiomic changes at multiple time points during treatment would also provide valuable insights into the dynamic nature of tumor response. Last, our analysis was restricted to T2-weighted images, and not all patients underwent all imaging sequences, such as contrast-enhanced MRI. Consequently, we cannot directly compare the efficacy of different imaging modalities. Finally, the study is limited by its exploratory mechanistic interpretation. The lack of paired histopathological and molecular data impedes a mechanistic explanation for the selected delta-radiomic features. Consequently, the linkage between imaging changes and treatment response remains inferential rather than causal.

In summary, T2-based MRI delta-radiomics represents a promising avenue for predicting chemotherapy response in osteosarcoma. By capturing and quantifying changes in tumor characteristics, this approach holds the potential to transform clinical practice and enhance personalized treatment strategies. Further validation and exploration of this technique could significantly impact the management of osteosarcoma and similar malignancies.

Author contributions

PY designed the study, collected and analyzed the data, and prepared and edited the paper. FZ collected and analyzed the data, and prepared and edited the paper. YL participated in the clinical research process and edited the paper. SS revised the paper. LL ensured the integrity of the whole study. HY participated in the clinical data acquisition and analysis. PFG collected and analyzed the data. JLR took part in the processing of data analysis and statistics. BSZ revised the paper. LHS revised the paper. NH designed the study, ensured the integrity of the whole study and revised the paper. All authors read and approved the final manuscript.

Funding

This study was supported by the National Natural Science Foundation of China (NO.82001764), Peking University People’s Hospital Scientific Research Development Funds (RDY2020-08, RS2021-10), and Beijing United Imaging Research Institute of Intelligent Imaging Foundation (CRIBJQY202105).

Data availability

The datasets used and/or analysed during the current study are available from the corresponding author on reasonable request.

Declarations

Ethics approval and consent to participate

The study adhered to the principles of the Declaration of Helsinki. The study was approved by the local ethics committee of Peking University People’s Hospital, and informed consent was waived due to the retrospective nature of the study.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.