Non-invasive MRI biomarkers for assessment of neoadjuvant chemotherapy response in osteosarcoma: Current techniques and clinical perspectives

Highlights

• •Advanced MRI allows early, noninvasive chemo response prediction in osteosarcoma, beating size-based criteria that miss tumor viability.
• •Diffusion & perfusion MRI are top modalities; ADC rise & DCE-MRI rWIR consistently link to histologic response.
• •rWIR from DCE-MRI is the best-validated biomarker, showing reproducibility & ties to event-free survival in multicenter studies.
• •Radiomics & radiogenomics show promise but are immature due to small samples, heterogeneity, & no external validation.
• •Interim MRI aids risk stratification & monitoring, but lacks proof to justify chemo changes without prospective validation.

Abstract

Background: Early prediction of chemotherapy response in osteosarcoma is critical for risk stratification, as the reference standard of histologic necrosis is only available after surgery. Conventional size-based imaging criteria are insufficient, prompting interest in advanced MRI biomarkers. To summarize and critically appraise the evidence on advanced MRI techniques for early prediction of response to neoadjuvant chemotherapy in high-grade osteosarcoma.

Methods: A narrative review of studies evaluating diffusion MRI, perfusion MRI (DCE-MRI), oxygenation-sensitive MRI (BOLD/R2*), radiomics, radiogenomics, and complexity analyses during or before neoadjuvant chemotherapy. Associations with histologic response and survival outcomes were synthesized qualitatively.

Results: Responders consistently demonstrated increased diffusivity (ADC/IVIM-D) and reduced perfusion during treatment. Among modalities, DCE-MRI relative wash-in rate (rWIR) showed the strongest and most reproducible association with histologic response and, in some cohorts, event-free survival. Radiomics and multiparametric models improved discrimination but lacked external validation. Performance depended on imaging timing, protocol standardization, and segmentation strategy.

Conclusion: Advanced MRI provides valuable early, noninvasive indicators of chemotherapy response in osteosarcoma, particularly diffusion and perfusion metrics. Currently, these techniques support risk stratification and monitoring rather than treatment modification. Prospective multicenter validation and standardized protocols are required for clinical adoption.

1. Introduction

Osteosarcoma is a high-grade primary bone malignancy in adolescents and young adults that is typically managed with neoadjuvant chemotherapy (NAC), definitive resection, and adjuvant therapy [1]. Despite contemporary protocols, survival has plateaued, and chemoresistance remains a major barrier, reinforcing the need for reliable early markers of treatment response [2]. As pathologic tumor necrosis, the current reference standard, is known only after surgery, clinicians lack robust tools to adapt to therapy during NAC.

Conventional magnetic resonance imaging (MRI) is essential for local staging and surgical planning. However, size-based criteria (for example, RECIST 1.1) correlate imperfectly with histological responses and can misclassify viable disease during or after NAC. This limitation has driven a shift toward functional and texture-based MRI readouts that better reflect tumor biology than gross dimensional changes [3], [4].

Evidence supports this direction, but warrants caution. Functional MRI approaches show predictive value, often after treatment has begun, which limits early decision-making [5], [6]. In parallel, the computational analysis of pretreatment MRI, capturing heterogeneity and morphological complexity, has reported encouraging discrimination. For example, monofractal/multifractal features classified chemo-responsiveness with bootstrap-corrected AUCs up to 0.88 [4]. These signals are associative and require validation; however, they illustrate the potential of texture-centric biomarkers [7].

Radiomics and radiogenomics extend this paradigm by quantifying textural features and linking them to the molecular context [8]. In a single-center cohort (n = 21), T2-weighted NGTDM “busyness” differentiated responders (AUC = 0.79) showed statistically significant correlations with genes related to resistance and immunity (GSTP1, CD274/PD-L1, and CCND3) [9]. The findings are preliminary, given the sample size and retrospective design; however, they outline a testable framework for noninvasive stratification.

This narrative review synthesizes current evidence on the use of advanced MRI for the early prediction of chemotherapy response in conventional high-grade osteosarcoma. We focus on functional sequences and computational methods, including radiomics and fractal analysis, summarize their associations with histologic necrosis and emerging clinical outcomes, and critically appraise their strengths, limitations, and areas of disagreement. Given the cross-center protocol variability, standardized acquisition/analysis and external validation are prerequisites for clinical adoption.

2. Methods

We conducted a narrative review of MRI-based biomarkers for early prediction of neoadjuvant chemotherapy response in conventional high-grade osteosarcoma. Modalities of interest included diffusion (DWI/ADC, IVIM, and DKI), perfusion (DCE-MRI; semi-quantitative and pharmacokinetic), T1 mapping, radiomics/radiogenomics, and complexity metrics. Given the heterogeneity and small sample sizes across studies, conclusions were framed as associations rather than causation and calibrated to study design and quality.

We searched PubMed/MEDLINE, Embase, Web of Science, and Google Scholar from inception to 1 July 2025, combining controlled vocabulary and keywords for osteosarcoma, MRI modalities, radiomics/radiogenomics, texture/fractal analysis, response/necrosis, and survival; reference lists of included papers and relevant reviews were hand-searched.

Eligible studies enrolled patients with pathologically confirmed osteosarcoma, evaluated baseline and/or interim MRI during NAC, and reported associations with histologic response, radiologic response, or survival (EFS/OS). We excluded case reports (<10 patients), conference abstracts without full text, non-MRI imaging only, purely technical/phantom work, and studies that lacked response endpoints. Two reviewers screened titles, abstracts, and full texts with arbitration by a third reviewer; overlapping cohorts were consolidated to the most complete or methodologically robust report.

The following cohort characteristics were extracted from each study: NAC regimen, MRI vendor/field strength, sequence parameters (DWI b-values, DCE temporal resolution, and AIF handling), segmentation approach (manual/semiautomatic; whole-tumor/whole-slab vs. focal), feature families, model type and validation, reference standards, and performance metrics (AUC, sensitivity/specificity, and calibration). For radiogenomics, we captured gene panels, MRI, and gene correlations. The risk of bias and reporting quality were appraised using tools appropriate for the study type (for example, QUIPS for prognostic factors, PROBAST for prediction models, and Radiomics Quality Score), feature stability, risk of leakage, inter-reader agreement, and the presence/absence of external validation [10]. Owing to protocol and endpoint heterogeneity, we performed a structured narrative synthesis by modality, reported effect directions, and performance ranges rather than pooled estimates and explicitly flagged conflicting results.

2.1. Study characteristics

Most of the included cohorts were retrospective, single-center, and small-to-moderate in size (typical n = 20–50). The participants were predominantly adolescents and young adults with conventional high-grade intramedullary osteosarcoma of the long bones; the femur was the most common, and neoadjuvant MAP-based regimens were typical [11].

The imaging modalities and time points varied. Studies have evaluated pre-treatment T2-weighted radiomics/radiogenomics [9], [12]; diffusion sequences before and/or after chemotherapy, including diffusion-weighted imaging (DWI) with apparent diffusion coefficient (ADC) [6], [13], [14], [15], [16], [17], [18], [19], [20], [21], [22], [23],intravoxel incoherent motion (IVIM) [3], [5], [19], [24],diffusion kurtosis imaging (DKI) [19], [25], perfusion with pre- and post-neoadjuvant DCE-MRI [5], [18], [26], and, in a minority, oxygenation-sensitive blood-oxygen-level-dependent (BOLD) MRI (R2* mapping) [27], [28]. Acquisition parameters and field strength were heterogeneous (e.g., earlier 1.5 T diffusion vs later 3 T T2-radiomics). Segmentation ranged from manual whole-tumor volumes to slice-based “whole-slab” or focal regions of interest [4], [20], [25]; inter-reader agreement for relative DCE features was high in a two-center study (acceptable intraclass correlation coefficients) [29].

Primary endpoints centered on the association with histologic response after resection (commonly ≥ 90% necrosis); survival endpoints were infrequently reported. Radiogenomics typically quantifies approximately 100 MRI texture features alongside targeted gene panels at baseline [12], [30]. However, external validation is uncommon.

2.2. MRI indicators of tumor response versus resistance

Across studies, functional change, not size, best mirrors histological response [Fig. 1]. Responders typically show increasing diffusivity (apparent diffusion coefficient (ADC) or mean diffusivity (MD), often paralleled by the IVIM true-diffusion coefficient D) and declining early enhancement/perfusion; in contrast, persistently restricted diffusion and brisk, sustained early enhancement suggest resistance [9], [31], [32]. Apparent volume enlargement may reflect treatment-related necrosis or cystic change; therefore, size-based criteria (e.g., RECIST) are insufficient in osteosarcoma [33].

Fig. 1.

Imaging evaluation of a 13-year-old male with high-grade conventional chondroblastic osteosarcoma of the distal femur. The patient underwent neoadjuvant chemotherapy with methotrexate, doxorubicin, and cisplatin (MAP regimen) and was considered for surgical resection after three courses. Baseline (A-D): (A, B) Anteroposterior and lateral radiographs demonstrating the primary lesion; (C) coronal T2-weighted STIR MRI; (D) axial T2-weighted MRI. Post-chemotherapy (E-H): (E, F) Anteroposterior and lateral radiographs of the distal femur after 12 weeks of treatment; (G) coronal T2-weighted STIR MRI; (H) axial T2-weighted STIR MRI.Histopathological analysis following resection revealed a poor response to chemotherapy, with approximately 20% tumor necrosis, corresponding to Huvos grade I.

Diffusion MRI (DWI/IVIM/DKI): Baseline mono-exponential ADC provides only modest discrimination (typical thresholds 1.3–1.4 × 10⁻³ mm²/s; AUC = 0.60–0.68 with mid-70% sensitivity/specificity) and does not generalize across scanners or b-value schemes [17], [34]. Post-treatment ADC sometimes separates groups better, but remains inconsistent, likely due to protocol and segmentation heterogeneity [28], [35]. Interim change after one chemotherapy cycle is more informative than baseline alone: failure of ADC, or the IVIM D parameter, to rise meaningfully tends to “rule in” poor response, although sensitivity varies with timing and region-of-interest (ROI) strategy [25], [32].

Reporting both absolute and percentage changes helps mitigate baseline variability [5]. IVIM adds a nuance: D usually mirrors ADC (higher in responders), whereas the pseudo-diffusion coefficient D* and perfusion fraction f often decrease with effective therapy [5], [36], [37]. Diffusion kurtosis imaging (DKI) strengthens microstructural readouts after neoadjuvant chemotherapy (NAC); in small cohorts, post-NAC MD outperformed ADC and mean kurtosis (MK) (AUC 0.91 vs 0.80 and 0.72) [33].

Practical considerations include avoiding universal cut-offs, favoring whole-tumor/whole-slab segmentation, masking cystic/necrotic regions to prevent spurious ADC inflation, and, where biologically relevant, analyzing the tumor core versus peri-tumoral rim separately [38], [39], [40]. Susceptibility, motion, and low-b noise particularly affect IVIM and DKI; thus, the explicit reporting of b-values, distortion control, and inter-reader reliability is essential [18], [21], [28].

Perfusion MRI (DCE-MRI): Early fast-enhancement metrics generally decreased in responders after one–two cycles and remained higher in non-responders [41]. The most mature evidence comes from a cohort study using whole-slab segmentation and a single threshold for relative wash-in rate (rWIR) < 2.3, yielding a cross-validated AUC of 0.93 (training) and 0.80 (external test), with inter-observer intraclass correlation coefficients (ICCs) of 0.81–0.97 for relative features, which is superior to that of focal ROIs [29].

Reproducibility was segmentation- and site-dependent, and the cross-center accuracy diminished without harmonization. Transparent reporting should include temporal resolution, precise feature definitions (e.g., wash-in rate (WIR), time-to-enhancement, maximum rate of enhancement (MRE), area under curve (AUC)), interval from the last chemotherapy to the scan, and ROI strategy (preferably whole-slab) [42].

Although these studies emphasized semi-quantitative metrics over pharmacokinetic modeling for workflow simplicity, the directionality is consistent and currently represents the most externally validated perfusion signal in the literature.

Oxygenation-sensitive MRI (BOLD-MRI). Blood-oxygen-level-dependent (BOLD) R2* mapping, an endogenous surrogate of deoxyhemoglobin, was feasible and reproducible [43]. In a small prospective cohort scanned pre-, mid-, and preoperatively, the change in R2* (ΔR2*) correlated more strongly with final necrosis, particularly in the extraosseous component, than with contemporaneous changes in k_ep or ADC [27]. No AUCs were reported and the power was limited; however, the consistent directionality and non-contrast nature make R2* a promising complementary marker for multiparametric models. Potential confounders include motion and field inhomogeneity, which should be acknowledged and mitigated whenever possible [44].

Radiomics and radiogenomics: Radiomics can contribute incremental prognostic signals beyond clinical variables and single-sequence metrics. In a study DWI/T2 clinical-radiomics nomogram outperformed clinical features and ADC alone (test-set AUC = 0.77–0.85) [21]. Multimodal models (e.g., combining radiography with MRI radiomics) reported AUCs of up to 0.83, although pairwise differences were not always significant, underscoring the overfitting risk [45]. Early radiogenomic studies linked T2-weighted texture (e.g., NGTDM “busyness”) with response and biologically plausible genes (GSTP1, CD274/PD-L1, HLA-I, and CCND3); however, the cohorts remained small [3], [12].

Complexity/fractal features on pretreatment MRI showed encouraging discrimination (best AUC = 0.88); however, they were sensitive to ROI handling and mono- versus multifractal choices [4]. Radiomics and radiogenomics should enhance, rather than replace, established functional readouts until larger, standardized, and externally validated studies confirm their reliability. While these MRI indicators hold promise for risk stratification and response monitoring, altering treatment based solely on these indicators has not yet been proven to improve outcomes. Consequently, claims regarding their routine integration should be approached with caution, emphasizing reproducible acquisition, pre-specified analysis, and prospective validation.

2.3. Timing of interim MRI and clinical utility

Early interim MRI aims to capture biological changes before measurable shrinkage, which is why size-based criteria (e.g., RECIST) underperform during neoadjuvant chemotherapy (NAC). In sarcomas, including osteosarcoma, apparent enlargement from necrosis, edema, or cystic changes can mimic progression; therefore, functional sequences are preferred for early readouts [40]. This rationale underpins mid-therapy diffusion and perfusion studies, and aligns with broader sarcoma guidance to avoid RECIST-only monitoring during this period [46].

Baidya Kayal et al. revealed that the strongest early signal typically emerges after the first cycle of chemotherapy. In a prospective intravoxel incoherent motion (IVIM) cohort scanned at baseline (t0), post–cycle 1 (t1), and completion (t2), discrimination for poor response improved from AUC 0.87 at t0 to 0.96 at t1 (sensitivity 86%, specificity up to 100%). Directionally, responders showed increasing diffusion (apparent diffusion coefficient (ADC) or IVIM true diffusion (D))and declining perfusion surrogates (D* and f, respectively).

Dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) studies have emphasized pre- versus post-NAC changes. However, real-world timing varies, and the relative wash-in rate (rWIR) has been operationalized before, after, or during the second cycle, supporting flexibility across common clinical time points [15]. Oxygenation-sensitive blood-oxygen-level-dependent (BOLD) MRI acquired pre-, mid-, and preoperatively showed statistically significant R2* increases over treatment, with ΔR2* (baseline → preoperative) correlating more strongly with final necrosis, particularly in the extraosseous component, than with contemporaneous changes in k_ep or ADC.

Although the most pronounced BOLD signal in that series reflected a longer interval, its noncontrast nature suggests a complementary value in multiparametric models [27], [28].

Thus, actionability remains a concern. A mid-course diffusion rule (ADC increase 0.07 × 10⁻³ mm²/s) achieved 100% specificity but only 25% sensitivity for identifying poor responders, useful to “rule in” high-risk biology, but insufficient alone to justify therapy change [6]. Similarly, the externally validated DCE model (whole-slab rWIR < 2.3) separates poor versus good histologic responders before surgery; however, the imaging response is not currently used to modulate NAC, and the accuracy thresholds for regimen changes remain unclear [29].

Performance also shifted with the interval from chemotherapy to scanning and with cross-center acquisition differences, underscoring the timing sensitivity and need for harmonization [30].

Therefore, a pragmatic framework is as follows: obtaining a high-quality baseline MRI for staging and segmentation, performing an early interim scan after one cycle to capture diffusion/perfusion shifts (the highest incremental yield in prospective IVIM data), and repeating at the end of NAC (preoperative) to inform resection planning and prognostication.

Treat-risk flags, minimal ADC/D rise, persistently high early enhancement, or rWIR < 2.3, as prompts for closer surveillance, multidisciplinary team review, or consideration of earlier surgery if other evidence of progression accrues, rather than as standalone triggers for switching chemotherapy. For prognostication, DCE-derived rWIR (externally validated and associated with event-free survival in some cohorts) and peritumoral perfusion/diffusion trends can be considered complementary [18]. These operational details will be discussed in the future [Fig. 2].

Fig. 2.

Radiological features in a 20-year-old man with conventional high-grade osteoblastic chondrosarcoma affecting the distal tibia. The patient received neoadjuvant chemotherapy and was later assessed for surgical excision, including reconstruction with a distal tibial allograft and tibio-talo-calcaneal arthrodesis. Baseline imaging at presentation (A-D): (A, B) Anteroposterior and lateral ankle radiographs illustrating the initial tumor; (C, D) axial and coronal T2-weighted STIR MR images. Following three cycles of chemotherapy (over 11 weeks) (E-H): (E, F) Anteroposterior and lateral radiographs post-treatment; (G, H) axial and coronal T2-weighted STIR MR images. Histologic evaluation after resection demonstrated a limited response to chemotherapy, showing approximately 30% tumor necrosis, consistent with Huvos grade I.

2.4. Association with survival outcomes

Survival endpoints are reported far less often than histologic responses and, where available, are derived mostly from small, single-center, retrospective datasets. Consequently, observed relationships should be interpreted as associations rather than causation [47]. Where links exist, they center on perfusion metrics from dynamic contrast–enhanced MRI (DCE-MRI) and, to a lesser extent, diffusion features; radiomics studies rarely include time-to-event analyses [22].

About DCE-MRI (rWIR) In a cohort (n = 82), the rWIR, using the same pre-specified cutoff as for histologic response, was independently associated with event-free survival (EFS) after adjustment for standard prognostic factors [29]. Patients with rWIR ≥ 2.3 had 2- and 5-year EFS rates of 85% and 75%, respectively, versus 55% and 50% for rWIR < 2.3; this pattern persisted among patients without metastases at diagnosis. In multivariable Cox models, rWIR < 2.3 was associated with a hazard ratio (HR) of 2.3 for shorter EFS. These data represent the strongest survival signals reported for MRI biomarkers in osteosarcoma, albeit retrospectively. Other perfusion parameters included a smaller series that suggested a complementary prognostic value from microvascular metrics. Lower post-chemotherapy K^trans (intratumoral) and lower baseline V_e (peritumoral) were associated with longer EFS and overall survival (OS) in one cohort (EFS, p = 0.002 and 0.026; OS, p = 0.003 and 0.023, respectively). These findings are hypothesis generating and warrant further prospective studies [41].

However, the association between Diffusion and survival is less consistent. In the same cohort, a higher post-chemotherapy peritumoral apparent diffusion coefficient (ADC) was correlated with a longer OS, whereas the tumor-core ADC results were mixed, likely reflecting variations in scan timing, region-of-interest strategy, and handling of heterogeneity/necrosis [15], [18], [28]. Oxygenation-sensitive BOLD MRI (R2*) has shown feasibility and stronger correlations with necrosis than ADC or DCE in a small prospective study; however, the survival endpoints have not been reported [27].

Radiomics/radiogenomics models primarily target histologic responses; durable links to EFS/OS remain largely untested and require standardized pipelines with external validation [9].

Among the MRI biomarkers, perfusion change, particularly rWIR, showed the clearest association with EFS, with smaller studies indicating the potential complementary value of K^trans/V_e and peritumoral ADC [8]. Given the retrospective design, modest sample sizes, and protocol heterogeneity, these signals should inform risk discussions and surveillance intensity rather than chemotherapy changes until prospective response-adapted trials demonstrate outcome benefits [48], [49]. Table 1 summarizes the main advanced MRI techniques investigated for early prediction of chemotherapy response in osteosarcoma, outlining their key biomarkers, biological interpretation, reported performance, and levels of validation across published studies [Table 1].

Table 1.

Summary of Advanced MRI Techniques for Early Prediction of Chemotherapy Response in Osteosarcoma.

MRI Technique	Key Parameters / Biomarkers	Findings / Biological Interpretation	Reported Evidence & Performance	Level of Validation / Limitations	Key References
Diffusion-Weighted Imaging (DWI)	Apparent Diffusion Coefficient (ADC)	↑ADC in responders (reduced cellularity & membrane breakdown); persistently low ADC → poor response	Baseline ADC modest (AUC 0.60–0.68); interim ΔADC after 1 cycle improves discrimination; thresholds ∼ 1.3–1.4 × 10⁻3 mm2/s	Widely available; protocol-dependent (b-values, ROI strategy); limited reproducibility across scanners	6, 13–23, 31, 34
Intravoxel Incoherent Motion (IVIM)	D (true diffusion), D* (pseudo-diffusion), f (perfusion fraction)	Responders: ↑D, ↓D* and f due to vascular pruning after NAC	Prospective data (AUC up to 0.96 post–cycle 1); D mirrors ADC trends; f and D* decline in good responders	Promising; requires high SNR and consistent b-values; sensitive to motion and noise	3, 5, 19, 24, 32, 36, 37
Diffusion Kurtosis Imaging (DKI)	Mean Diffusivity (MD), Mean Kurtosis (MK)	Captures non-Gaussian diffusion; reduced MK and increased MD indicate treatment response	Post-NAC MD AUC 0.91 (better than ADC: 0.80, MK: 0.72) in small cohort (n < 30)	High microstructural sensitivity; technically demanding; small samples	19, 25, 33
Dynamic Contrast-Enhanced MRI (DCE-MRI)	Relative Wash-In Rate (rWIR), Ktrans, Ve, Kep	↓Early enhancement and ↓rWIR after effective NAC (microvascular collapse); high rWIR = poor response	rWIR < 2.3 → AUC 0.93 (training) / 0.80 (external test); correlated with EFS (HR 2.3 for rWIR < 2.3)	Most validated MRI biomarker; requires contrast and temporal standardization	5, 18, 22, 26, 29, 41
Blood-Oxygen-Level-Dependent MRI (BOLD-MRI)	R2* (ΔR2*)	↑R2* change indicates improved oxygenation and necrosis; correlates with histologic necrosis	Prospective study: ΔR2* correlated with necrosis better than ADC or kep (no AUC reported)	Non-contrast, reproducible; limited by motion and field inhomogeneity; small cohorts	27, 28, 43, 44
Radiomics / Radiogenomics	Texture features (GLCM entropy, NGTDM “busyness”), gene correlations (GSTP1, PD-L1, CCND3)	Pretreatment heterogeneity (high entropy/busyness) → poor response; linked to chemoresistance genes	AUC 0.77–0.85 (radiomics models); T2 “busyness” AUC 0.79; GSTP1 & PD-L1 correlation confirmed	Promising but small single-center studies; no external validation; overfitting risk	3, 9, 12, 21, 30, 45, 49
Fractal / Complexity Analysis	Fractal dimension, coefficient of variation λ′(G)	Higher morphological complexity and irregularity → chemoresistance	Monofractal CV(λ′(G)) AUC 0.88 (bootstrap-corrected)	Conceptually robust to ROI variation; needs prospective, standardized validation	4, 7, 59
Multiparametric / AI-based Models	Combined DWI + DCE + Radiomics	Integration improves discrimination and prognostic accuracy	Combined nomograms AUC up to 0.85–0.90 (internal test); improved calibration	Requires harmonized datasets, external validation, and explainability	21, 38, 45, 49

3. Discussion

This narrative review consolidates the evidence that advanced MRI techniques can provide early, noninvasive indicators of chemotherapy response in osteosarcoma.

Functional MRI modalities, diffusion-weighted imaging, dynamic contrast-enhanced MRI, intravoxel incoherent motion, and computational approaches, such as radiomics and fractal analysis, detect microstructural and microvascular changes associated with tumor necrosis before gross volumetric changes occur. Across studies, these techniques have demonstrated promising predictive accuracy in distinguishing between good and poor responders, outperforming the conventional size-based criteria. However, methodological heterogeneity in the literature (e.g., variable imaging protocols, sample sizes, timing, and endpoints) underscores the need for standardized acquisition and analysis to translate these MRI biomarkers into routine clinical practice [50].

Advanced MRI findings can be interpreted in the light of osteosarcoma biology. In responding tumors, effective chemotherapy reduces cellularity and disrupts cell membranes, allowing water to diffuse more freely, manifesting as increased ADC values on DWI [17], [18]. Empirically, post-neoadjuvant ADC increases are strongly correlated with histopathological necrosis, supporting diffusivity as a proxy for cell death [6], [12]. Conversely, persistently low ADC values indicate retained tumor cellularity and chemoresistance [51]. Similarly, DCE-MRI captures the collapse of the tumor microvasculature in good responders; metrics of early tumor perfusion/permeability (for example, K^trans and wash-in rate) decrease markedly after successful treatment [40].

In a two-center study, a pre-specified rWIR threshold discriminated histologic responders with high inter-observer agreement; full performance and reliability details are provided in the Results section (Perfusion MRI). Biologically, a decrease in K^trans or rWIR reflects reduced blood flow and vessel permeability as viable tumor tissue is ablated, whereas tumors that remain well-perfused despite chemotherapy tend to be non-responders [18], [52]. IVIM MRI adds nuances to this picture by decoupling pure diffusion (parameter D) from perfusion-related diffusion (D*). Studies have shown that after a few cycles of therapy, the D value increases in responders (mirroring ADC), whereas the perfusion fraction (f) and D* often decrease owing to pruning of the tumor microcirculation [32], [52], [53]. These physiological insights explain why functional MRI changes consistently track treatment effects, whereas conventional MRI tumor size changes (RECIST) often lag or misrepresent actual tumor viability.

Radiomics and radiogenomics extend beyond human-visible features by quantifying tumor heterogeneity and linking it to molecular profiles. Several MRI-based radiomics models report encouraging discrimination versus single-sequence metrics, but require external validation. for predicting histologic response, frequently outperforming single parameters such as ADC or K^trans. [21] These models capture texture patterns indicative of necrosis or viable tissues intermixed within the tumor. For example, high Gray-Level Co-occurrence Matrix (GLCM) entropy or Neighborhood Gray-Tone Difference Matrix “busyness” on pretreatment scans often signifies more disordered, heterogeneous tumors that tend to respond poorly [3], [9], [54]. The emerging field of radiogenomics has reinforced the biological relevance of these features.

In a pilot study, Yin et al. found that a T2-weighted MRI texture feature (NGTDM busyness) not only differentiated responders (AUC = 0.788) but also correlated with the expression of chemoresistance-related genes, such as GSTP1 and the cell-cycle regulator CCND3 [9]. GSTP1 encodes a glutathione S-transferase involved in the detoxification of chemotherapeutic agents, providing a mechanistic link between the imaging phenotype and drug resistance [55], [56]. Tumors with high intratumoral heterogeneity on MRI show elevated GSTP1 expression, potentially reflecting subregions that can neutralize chemotherapy [9]. Such integration of imaging with genomics suggests a precision oncology approach, where an MRI scan can noninvasively flag tumors with molecular profiles (e.g., high glutathione metabolism or immune checkpoint expression) that confer chemoresistance [57]. Despite promising developments in radiogenomic research, the findings remain preliminary owing to the limited sample size (n = 21) and the absence of external validation to date.

Computational fractal analysis is another novel modality that complements standard radiomics by quantifying tumor shape complexity and spatial scaling behavior. Djuričić et al. introduced fractal texture features derived from pretreatment MRI and reported that a monofractal metric (the coefficient of variation of the grayscale distribution function, CV for λ′(G)) achieved an AUC of 0.88 for predicting poor histologic response [4]. Lower tumor homogeneity and higher morphological irregularity on MRI are strongly associated with chemoresistant osteosarcoma [28]. Fractal features capture the scale-invariant heterogeneity that conventional metrics may miss; for example, they can reflect the complex and chaotic architecture of an aggressive tumor [58].

This approach showed robustness to ROI delineation, a valuable trait, because inconsistent tumor segmentation can hamper radiomic reliability [4], [59]. However, fractal analysis currently requires specialized software (e.g., FracLac) and expertise, and like radiomics, it must be validated prospectively. Notably, both the radiogenomic and fractal studies were retrospective and relatively small, so their impressive performances could be optimistic without further validation (the fractal study did use internal bootstrapping to mitigate overfitting risk).

4. Clinical Implications

Together, these advanced MRI modalities may enhance early response assessment and risk stratification, supporting adaptive monitoring strategies, rather than treatment changes alone. Functional MRI changes appear to be more reliable indicators of response than gross tumor size, which often correlates modestly with necrosis. If validated, MRI-based response assessment could inform risk stratification and multidisciplinary planning; however, whether altering therapy based on this would improve outcomes requires prospective response-adapted trials.

For instance, a patient whose mid-therapy MRI shows persistently low ADC and sustained high perfusion (suggesting poor response) should prompt closer surveillance, multidisciplinary review, and consideration of trial enrollment [40]. Early surgery should be performed only if it is corroborated by additional evidence of progression [8], [60]. Conversely, a patient with a markedly rising ADC or declining K^trans after one or two cycles (indicating a robust response) may safely continue standard therapy and avoid premature treatment escalation or unnecessary toxicity [61].

Such response-adaptive decisions are not possible with current practices, where a definitive histologic response is only known after resection. In an era of platuized survival rates (60–70% 5-year survival in localized osteosarcoma), early identification of non-responders is crucial and could identify candidates for clinical trials or intensified monitoring; however, whether early regimen switches improve outcomes remains unproven [32], [62]. This may potentially improve outcomes for 20–30% of patients who otherwise endure ineffective first-line chemotherapy. Importantly, MRI-based biomarkers are radiation-free and relatively cost-effective, which is advantageous over FDG-PET, which has been explored for response assessment but has a higher cost and radiation exposure [63].

Functional MRI changes (such as IVIM-DWI and ADC) parallel the information from metabolic PET (tumor cell kill and necrosis), but with greater availability in routine care [64], [65]. Thus, integrating advanced MRI into osteosarcoma pathways could support risk stratification and inform surgical planning discussions [38]. Decisions regarding adjunct therapies await prospective validation, aligning with the personalized treatment principles emphasized in modern oncology guidelines.

5. Limitations and challenges

Despite these encouraging results, several challenges must be addressed before advanced MRI biomarkers can be clinically adopted. A major limitation of this review is the heterogeneity of existing studies. Most cohorts are single-institution and modest in size (median N = 25–50), often mixing pediatric and adult patients or various tumor locations, which confounds their generalizability. Many studies have excluded axial or pelvic osteosarcoma to focus on extremity tumors, improving internal consistency at the cost of real-world applicability.

There is also inconsistency in chemotherapy protocols and the timing of post-therapy imaging across studies, which complicates the comparisons. Technical variability is another issue: MRI field strength (1.5 T vs. 3 T), pulse sequence parameters (for example, DWI b-values ranging from 800 to 2000 s/mm²), and definition of endpoints (e.g., ≥90% vs. ≥ 95% necrosis as “good response”) differ widely. This lack of standardization can significantly affect the absolute values of imaging metrics and their predictive performance. For example, an ADC threshold that separates responders in one center might not be applicable to another if diffusion protocols or machines differ [35].

Harmonization techniques, such as ComBat normalization, have shown that multi-center radiomic models maintain accuracy when scanner differences are adjusted, whereas performance can drop if such effects are not mitigated [29], [30]. Going forward, community consensus on MRI acquisition (e.g., standardized DWI b-values and DCE temporal resolution) and analysis pipelines is essential for reducing site-to-site variability.

Another concern is the potential for overfitting and statistical bias. Radiomics involves mining hundreds of features from relatively small datasets, raising the “curse of dimensionality.” [66] Without rigorous feature reduction and validation, the models may simply fit the noise.

Many studies have employed techniques such as LASSO or principal component analysis to select features (for instance, Chen et al. used LASSO to build a multifeature MRI radiomics model), but only a few have provided independent external validation of their models [21], [54]. The paucity of external validation is notable: aside from the recent DCE-MRI model that was tested on a separate cohort [29], most predictive models remain unconfirmed outside their derivation set. This implies that their real-world performance is uncertain.

Moreover, multiple hypothesis testing is inherent in radiomic and genomic analyses; without proper correction, there is a risk that some reported significant predictors are false positives [30], [49]. Future studies should pre-specify the primary endpoints and adjust for multiple comparisons to ensure robust results [38].

Interobserver and technical reproducibility also pose challenges. Many MRI biomarkers require tumor segmentation, which can be performed manually and is prone to variability. Differences in how radiologists draw tumor ROIs can alter quantitative metrics such as ADC (especially if necrotic or cystic areas are inconsistently included). Encouragingly, when whole-tumor “slab” ROIs are used (covering the largest cross-section of the tumor), the interobserver agreement is high for DCE parameters (ICC = 0.81 to 0.97), which is much better than focusing on small hotspot ROIs (which had an ICC as low as 0.57) [29].

However, the widespread use of these techniques requires training and standardized protocols for ROI delineation. Automated segmentation algorithms, possibly using deep learning, can alleviate this issue by providing consistent tumor volumes for analysis [38], [49]. However, these methods require further validation and regulatory approval.

From an implementation standpoint, integrating radiomics and AI-driven analyses into clinical workflows is nontrivial. Many radiology departments lack the infrastructure and expertise required for advanced image postprocessing. Radiomics and radiogenomic analyses often require custom software or coding experience, meaning that community hospitals might find it difficult to reproduce the results reported by academic centers. There are also regulatory hurdles: if an algorithm is developed to predict the response (for example, a machine learning model combining ADC and radiomic features), it may be considered a medical device and thus requires regulatory clearance (e.g., FDA approval in the US) before clinical use.

Additionally, standardizing such algorithms for deployment across different MRI vendors and software platforms is a challenge that needs to be coordinated by professional societies or multicenter consortia. Finally, clinical acceptance and ethical aspects must be considered. Oncologists and radiologists should be convinced of the added value of MRI biomarkers and trained to interpret them. There may be hesitancy to change or escalate therapy based on an imaging predictor alone without histological confirmation; thus, a clear demonstration of patient benefits will be key.

Ethically, the use of an early predictive biomarker raises questions, such as how to communicate with patients if an MRI suggests a poor response after only a few weeks of therapy (potentially causing anxiety) and how to ensure that access to such advanced imaging does not widen healthcare disparities. Fortunately, MRI is widely available relative to other modalities such as PET; therefore, an MRI-based approach could be accessible even in resource-limited settings, provided that the techniques are cost-effective and user-friendly. Ensuring equitable implementation (through cost reduction, training programs, and possibly AI tools that can be deployed in standard hospital PACS systems) will be important so that all patients can benefit from these advances in the future.

6. Future directions

The next phase requires preregistered multicenter studies with harmonized DWI/DCE protocols, standardized post-processing, and embedded external validation. Within-patient head-to-head comparisons should quantify the incremental value of diffusion (ADC/IVIM), perfusion (including rWIR), oxygenation-sensitive mapping, and radiomic and fractal features. Multiparametric models that integrate imaging with clinical variables and molecular biomarkers (e.g., circulating tumor DNA) [67] should be developed with transparent feature selection, rigorous calibration, and decision curve analysis, and tested against patient-centered endpoints (EFS/OS) using pre-specified action thresholds.

Closing the implementation gap will depend on reproducible, scalable pipelines, automated segmentation, cross-scanner harmonization, vendor-neutral software, and clear reporting standards, as well as response-adapted trials that randomize management based on early interim MRI to prove the outcome benefit. Parallel usability, regulatory, and health economic evaluations are needed to ensure equitable adoption beyond tertiary centers. Together, these steps will determine whether MRI biomarkers can qualify as surrogate endpoints and be integrated into clinical pathways if the trials demonstrate benefits.

7. Conclusion

In conclusion, advanced MRI techniques show promise as early indicators that complement current assessments by providing insights into chemotherapy efficacy that were previously obtainable only after surgery. These functional and computational MRI biomarkers, from ADC and IVIM parameters to radiomic textures, radiogenomic signatures, and fractal dimensions, capture the fundamental aspects of tumor biology (cellularity, perfusion, heterogeneity, and molecular activity) that underlie the treatment response. Although early validation studies are encouraging, rigorous prospective evaluations are necessary to fully establish their clinical utility. With continued research and standardization, MRI-based markers could eventually serve as reliable surrogate endpoints, informing risk discussions and follow-up. Whether response-adapted strategies improve outcomes in aggressive cancers remains unclear. Radiogenomics, particularly fractal analysis, represents a cutting-edge frontier that bridges imaging with genomics and mathematics to better stratify patients. Realizing this promise requires multidisciplinary collaboration and validation. However, the potential benefit of more personalized and responsive care for osteosarcoma is a critical direction for future research.

Declarations

Ethics approval and consent to participate

Not applicable. This article is a narrative review of previously published studies; no human participants were enrolled and no identifiable data were used.

Consent for publication

Not applicable.

Availability of data and materials

No new datasets were generated or analyzed. All data discussed are available in the cited literature. The full search strategy, screening log, and data-extraction template are provided in the supplementary material and are available from the corresponding author on reasonable request.

CRediT authorship contribution statement

Hamed Naghizadeh: Writing – review & editing, Writing – original draft, Validation, Methodology, Data curation, Conceptualization. Masih Rikhtehgar: Writing – review & editing, Formal analysis, Data curation, Conceptualization. Sadegh Saberi: Methodology, Data curation. Elham Poorghasem: Writing – review & editing, Writing – original draft, Data curation. Khodamorad Jamshidi: Supervision, Methodology, Investigation, Data curation. Seyyed Saeed Khabiri: Writing – review & editing, Writing – original draft, Methodology, Formal analysis, Data curation, Conceptualization.

Funding

This research received no specific grant from any funding agency in the public, commercial, or not-for-profit sectors.

Declaration of competing interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.