Measurable imaging-based changes in enhancement of intrahepatic cholangiocarcinoma after radiotherapy reflect physical mechanisms of response Authors: Brian De, Prashant Dogra, Mohamed Zaid, Dalia Elganainy, Kevin Sun, Ahmed M Amer, Charles Wang, Michael K Rooney, Enoch Chang, Hyunseon C Kang, Zhihui Wang, Priya Bhosale, Bruno C Odisio, Timothy E Newhook, Ching-Wei D Tzeng, Hop S Tran Cao, Yun S Chun, Jean-Nicholas Vauthey, Sunyoung S Lee, Ahmed Kaseb, Kanwal Raghav, Milind Javle, Bruce D Minsky, Sonal S Noticewala, Emma B Holliday, Grace L Smith, Albert C Koong, Prajnan Das, Vittorio Cristini, Ethan B Ludmir, Eugene J Koay Published: 2025 ABSTRACT AbstractEscalated doses of radiotherapy associate with improved local control and overall survival (OS) in intrahepatic cholangiocarcinoma (iCCA), but personalization remains limited because conventional size-based CT criteria correlate poorly with outcomes. We hypothesized that quantitative enhancement measurements would better predict clinical outcomes and guide individualized RT optimization. In a retrospective cohort of 154 patients, we analyzed pre- and post-RT CT scans using quantitative European Association for Study of Liver (qEASL) to derive viable tumor volumes, comparing enhancement-based metrics with size-based RECIST and linking them to outcomes via survival and mathematical modeling. Change in enhancement volume was strongly associated with OS after adjustment, outperforming RECIST, and a ≥ 33% reduction optimally distinguished responders. From modeling analyses, the patient-specific tumor growth rate parameter emerged as the dominant mechanistic predictor, achieving 80.5% classification accuracy. Importantly, CT-derived mathematical parameters from this framework may inform RT planning and dose adaptation, particularly for resistant tumors, by bridging imaging with mechanistic insight.Subject terms: Cancer, Computational biology and bioinformatics, Oncology Introduction Intrahepatic cholangiocarcinoma (iCCA) is an uncommon and aggressive malignancy arising from the epithelial lining of the intrahepatic biliary tree, for which 70% of patients present with unresectable tumors1. While chemotherapy has historically been considered the standard of care for these patients, treatment with chemotherapy alone has been associated with unsatisfactory results, with a median overall survival (OS) of 3–12 months2. Recent retrospective studies have shown that treatment with ablative doses of radiotherapy (RT) is associated with local control (LC) and OS in patients with iCCA3–5. However, the RT dose and fractionation schedule are subjectively chosen for each patient, and there is no consensus on the optimal protocol for treating iCCA3,6. Further compounding this uncertainty, evidence suggests that the optimal treatment of iCCA may be specific to individuals, potentially mediated by differences in pretreatments received, functional liver reserve, and intrinsic radiosensitivity7. To address these intricacies, investigators in the field of mathematical oncology8 have developed several mathematical models to predict RT response9, with the linear quadratic (LQ) model being the most prevalent. The LQ model posits that radiation-driven cell death results from double-strand DNA breaks caused directly by ionizing radiation, or through interactions involving single-strand DNA breaks10. Leder et al. formulated a more comprehensive model encompassing tumor cellular diversity and dynamically evolving radioresistance, enabling prediction of the efficacy of radiation schedules11. Prokopiou et al. introduced the proliferation saturation index (PSI), which implicitly links the proliferative fraction of a tumor with its radiation sensitivity12. Zahid et al. adapted the PSI model to simulate the impact of radiation not through LQ cell survival but by altering the tumor microenvironment, thereby enabling the forecasting of patient-specific responses to radiation13. To quantify radiosensitivity from a genomic perspective, modeling has been applied to develop a radiation sensitivity index (RSI) which connects the expression of specific genes to the radiosensitivity of cells14. Building on this, a genome-based model for adjusting radiotherapy dose (GARD) was developed by Scott et al., linking patient-specific RSI values and the LQ model with RT dose plans to customize treatment schedules based on patient radiosensitivity15. To enhance the clinical application of mathematical models, we previously integrated pre-treatment CT imaging with mechanistic models to predict clinical responses in pancreatic cancer. These models are instrumental in understanding cytotoxic response mechanisms and have the potential to inform therapeutic decisions16. CT-based imaging for iCCA prior to and following RT provides a rich source of data from which mathematical models may permit therapy selection and adaptation. Nevertheless, there are shortcomings in the current methods of radiographic evaluation of hepatic tumors. Currently, approved criteria for assessing tumor response to treatment include those from the World Health Organization (WHO) and Response Evaluation Criteria in Solid Tumors (RECIST)17–19. Both criteria depend on changes in overall tumor size using the largest unidimensional or bidimensional measurements. A significant limitation of these size-based criteria is that they do not account for tumor morphological changes that reflect treatment response, such as tumor necrosis and decreased tumor perfusion, which may occur after liver-directed RT20. Early prediction of treatment response after RT in iCCA may substantially impact further management, particularly if other liver-directed or systemic therapies are viable options. To address this, an expert panel introduced the European Association for the Study of the Liver (EASL) criteria for hepatocellular carcinoma (HCC), which considers the enhancing portion of the tumor as viable tissue using bidimensional measurements of the largest enhancing portion of the tumor21. The EASL criteria have been shown to be predictive of patient outcomes and superior to RECIST and WHO criteria in the assessment of response after treatment of HCC22–25. The utility of CT-based tumor enhancement for response assessment following RT for iCCA is yet to be explored and may provide more accurate outcome prediction than traditional size-based criteria26. More importantly, based on the dynamic assessment of viable tumor volume, adaptive RT can be supported by integration with mechanistic mathematical models of tumor growth and treatment response27. Therefore, the use of qEASL-based mathematical models may be an attractive means of personalizing RT treatment protocols. In the present study, we tested the hypothesis that changes in tumor enhancement following RT can inform a mathematical model of cytotoxic response in iCCA. We aimed to evaluate the utility of changes in tumor enhancement for RT response assessment in iCCA and predict outcomes following treatment. In addition, using qEASL-based viable tumor volume measurements, we sought to develop a mechanistic mathematical model as a prospective tool for personalization and adaptation of RT. An imaging-based biomarker of treatment response, integrated with mathematical modeling, may better guide the subsequent management of iCCA patients, which is important in both direct patient care and clinical trial settings. Results Patient characteristics Patient characteristics are shown in Table 1. The analytic cohort consisted of 154 patients. The median age was 64 years (range, 29–88 years), and 74 (48%) patients were female. The median RT dose was 62.5 Gy (range, 35–100 Gy), and the median number of fractions was 15 (range, 3–31 fractions). The median interval between the end of RT and post-RT CT scan was 1.8 months (range, 0.1–10.9 months). RT technique was proton in 37 patients and photon in 117 patients. Table 1Patient characteristics among all patients and stratified by qEASL responseAll patientsqEASL respondersqEASL non-respondersP-valueNumber1547282Age, median [range]64 [29–88]63 [29–86]65 [33–88]0.069Sex0.747 Female743638 Male803644Time interval between end of RT and post-RT CT scan (months), median [range]1.8 [0.1–10.9]1.9 [0.2–9.3]1.8 [0.1–10.9]0.051AJCC 8th edition stage0.385 I1275 II291613 III20614 IVA442222 IVB492128Tumor largest dimension (mm), median [range]72 [17–179]74 [17–179]72 [24–176]0.865Intrahepatic satellites at diagnosis0.202 Yes763145 No784137ECOG scale0.687 0522230 1904446 21165 3101Concurrent chemotherapy0.234 No522824 Yes1024458Radiation dose Gy median [range]62.5 [35–100]66 [50–100]60 [35–100]0.124BED ≥ 80.5 Gy0.069 Yes1135855 No411427Technique0.004* Photon1174770 Proton372512*Significant at 5% level. AJCC American Joint Committee on Cancer, BED Biologically effective dose, CT Computed tomography, Gy Gray, ECOG Eastern Cooperative Oncology Group, qEASL Quantitative European Association for the Study of Liver Criteria. Patient characteristics among all patients and stratified by qEASL response *Significant at 5% level. AJCC American Joint Committee on Cancer, BED Biologically effective dose, CT Computed tomography, Gy Gray, ECOG Eastern Cooperative Oncology Group, qEASL Quantitative European Association for the Study of Liver Criteria. Disease outcomes Patient outcomes following RT have been previously reported by our institution3,5. In the current cohort, updated with the last available follow-up, the median follow-up time was 64 months (range, 43–80 months) following RT. At latest follow-up, 63 patients (41%) had experienced local failure, and 127 (82%) had died. The median LC was 30 months (95% confidence interval [CI] 19–44 months), and median OS was 18 months (CI, 15–21 months). RECIST 1.1 assessment Patients were classified by RECIST 1.1 criteria as having complete response (CR; no visible tumor), partial response (PR; decrease of ≥30% in sum of longest diameter), progressive disease (PD; increase of ≥20% in sum of longest diameter), or stable disease (SD; all others)19. At the first follow-up scan after completion of RT, the median longest tumor diameter did not significantly change; the median was 7.2 cm (interquartile range [IQR], 5.2–10.3 cm) before RT and 6.8 cm (IQR, 4.9–-10.0 cm; P = 0.115) after RT. The training set had 8 patients with PR to RT, 86 patients with SD, 14 patients with PD, and no patients with CR. As shown in Fig. 2A, B, there was no significant difference in OS between the various response categories (P = 0.7052), but there was a significant difference in LC (P = 0.0137) with stratification by RECIST 1.1 response. Comparison of survival curves stratified by RECIST and qEASL-based evaluations. Overall survival (A) and local control (B) among patients stratified by response as measured by RECIST 1.1 and overall survival (C) and local control (D) among patients stratified by response as measured by qEASL. qEASL assessment Ground truth responder labels for all analyses were defined based on LC duration, with a threshold of 24 months used to distinguish responders (n = 34) and non-responders (n = 120). Baseline characteristics stratified by qEASL (responders and non-responders) are shown in Table 1. A key difference was greater use of proton therapy in the qEASL-responder cohort. As shown in Supplementary Table 2, univariate Cox analysis showed that a decrease in qEASL-derived viable volume was associated with prolonged OS (P < 0.001) but not LC (P = 0.083). Similarly, in multivariable analysis, a decrease in viable volume as a continuous variable was significantly associated with longer OS (P < 0.001) but not LC (P = 0.088) after adjusting for age, stage, and BED. The change in viable volume (%) was evaluated for its ability to distinguish responders from non-responders. Using binary classification, a cut-off point of 33% decrease in viable volume was determined to optimally predict prolonged LC ( > 24 months; Fig. 1E; area under the ROC curve [AUC] = 0.78). The corresponding sensitivity, specificity, and accuracy of the classifier were 85.3%, 65.0%, and 69.5%, respectively. Further, as shown in Fig. 1F, LOOCV upheld the validity of the classifier with an average AUC of 0.77±0.003 and sensitivity, specificity, and accuracy of predicting the left-out data point at 85.3%, 64.2%, and 68.8%, respectively. Furthermore, patients classified as responders according to the qEASL criteria demonstrated a significantly longer median OS (34 vs. 13 months; P < 0.0001) and median LC (44 vs. 15 months; P < 0.0001) (Fig. 2C, D). Binary classification of qEASL-based classifier. Representative CT images from (A, C) pre-RT and (B, D) post-RT in patients who were identified as (A, B) a qEASL responder and (C, D) a qEASL non-responder. E ROC curve to determine qEASL-based cutoff point for percentage change in enhancement volume following RT. Red circle on the ROC curve denotes the optimal operating point identified through the Jaccard index. F Cross-validation of the binary classifier. The corresponding confusion matrices are displayed below the figures. Molecular correlates of response to treatment Mutational frequencies and their associations with treatment responses are shown in Supplementary Table 3. The most commonly observed mutations were in TP53, IDH1, BAP1, and ARID1A. Univariate linear regression was performed to identify potential relationships between mutations and treatment response, as measured by the percentage change in enhancement volume following RT. Only one potential association was identified: ARID1A mutation was associated with a 281% increase in enhancement volume vs. a 13% decrease in enhancement volume in wild-type patients (P = 0.034). Modeling-derived predictors of radiotherapy response As shown in Fig. 3A, B, the model-fitted trajectory of patient tumor enhancement volume over time (orange line), under the patient-specific radiation treatment implemented in the simulation (blue line), demonstrates the difference in the evolution of tumor dynamics in a representative responder (patient ID 54; Fig. 3A) and non-responder (patient ID 47; Fig. 3B). The key model parameter values, estimated from CT imaging (i.e., BVF) and model fitting (i.e., γ and α), shown in the inset of the figures, suggest that in the specific examples, despite the same dose and fractions given to the two patients, the tumor response varied significantly owing to differences in tumor growth rate γ and BVF. A responder has a slower growth rate and higher BVF than a non-responder. As we know from the model equation, because of the difference in BVF, the D50eff in the two cases varied by an order of magnitude, such that the potency of radiation in a non-responder appears to be much lower than that in a responder. In addition, a larger value of γ obtained for a non-responder suggests a more aggressively growing tumor, indicating a net positive growth rate under treatment. Representative model fits and parameter comparisons between responders and non-responders. A–D Fitted tumor enhancement volume trajectories Venht (orange) and modeled residual radiation effect Dt (blue) for four cases: a representative responder (A, patient ID 54), non-responder (B, patient ID 47), and averaged responder (C) and non-responder (D) profiles. Orange circles denote raw enhancement volume data extracted from CT imaging. Parameter estimates for each case (tumor growth rate γ, resistance coefficient α, perfusion-derived BVF, effective inhibitory dose D50eff, total dose, and number of fractions) are shown in the inset. E Group-wise comparison of model-fitted and data-derived parameters for patients classified using local control (LC ≥ 24 months) as the ground truth criterion (R responders, n = 34; NR: non-responders, n = 120). Bars represent the group mean, and error bars indicate standard error of the mean. Statistical significance was assessed using Welch’s t-test. *P < 0.05; ***P < 0.0001; ns: not significant. Treatment gap refers to the average time interval (days) between radiation fractions. Furthermore, the model was fit individually to each patient (Supplementary Fig. 1), and parameter estimates were obtained (Fig. 3E, Supplementary Table 5). Using the average parameter values shown in Fig. 3E, we simulated the tumor response dynamics for an average responder and an average non-responder, starting with the average initial tumor enhancement volume across all patients, that is, ~25 cm3 and ~17 cm3, respectively (Fig. 3C, D). From these simulations, we observed that the average responder showed a ~ 51% reduction in enhancement, whereas the average non-responder exhibited only ~11% decrease in enhancement during the course of treatment. Extending the simulation beyond the end of treatment up to day 150 demonstrated that as the dose washes out, tumor enhancement eventually begins to increase, although at much different rates for the two groups (due to significant differences in their γ parameter values), such that the enhancement volumes at day 150 were 6 cm3 and 22.5 cm3 for responders and non-responders, respectively. The above observations can be contextualized from the parameter estimates shown in Fig. 3E. The tumor growth rate constant γ, radiation resistance coefficient α, effective half-maximal inhibitory radiation dose or radiation potency D50eff, and pre-RT viable volume were the only parameters that were significantly different between the responder and non-responder groups (Welch’s t-test, P < 0.05). This indicates that owing to relatively smaller growth rates, lesser radiation resistance, and higher radiation potency, the responder group showed superior treatment outcomes to the standard RT regimens. Note that while the pre-RT tumor volumes were significantly smaller in the responder cohort, the corresponding difference between the BVF values (which is the ratio of pre-RT tumor enhancement volume to pre-RT tumor volume) between the two groups was insignificant. This suggests that D50eff ( = D50BVFα) is predominantly affected by the desensitization coefficient α. To evaluate whether the fixed mechanistic parameters in the model (δmax, D50, and k) unduly influence tumor response predictions, we performed a Global Sensitivity Analysis (GSA) using Latin Hypercube Sampling and Spearman rank correlation (Supplementary Fig. 2). We simulated tumor enhancement volume dynamics by varying δmax, D50, and k within biologically plausible ranges while holding cohort-averaged values of γ, α, BVF, and baseline tumor enhancement volume fixed for responders and non-responders, respectively. The results show that δmax had the strongest (inverse) association with final tumor volume (ρ = –0.88 in responders, −0.86 in non-responders), consistent with its role in amplifying tumor death when radiation is effective. Importantly, its realized effect is always scaled by patient-specific parameters (γ, α, BVF) and dose schedule, preventing δmax from dominating outcomes on its own. The decay constant k also showed a strong positive influence (ρ = 0.45 in responders, 0.46 in non-responders), reflecting its modulation of the time course of the radiation effect D(t). D50 exhibited a moderate effect (ρ = 0.17 in responders, 0.23 in non-responders), driven by its interaction with BVF and α via the D50eff term. These findings demonstrate that all three parameters influence model output in distinct and interpretable ways, but that response heterogeneity primarily emerges from the patient-specific variables (γ, α, BVF) rather than arbitrary fixed values. This supports both the robustness of the model structure and the rationale for limiting free parameters to preserve identifiability under sparse imaging data. We next evaluated the predictive power of the four patient-specific parameters that differed significantly between responders and non-responders using logistic regression-based classification. As shown in Fig. 4A, the ROC curves obtained for the four classifiers (α, γ, D50eff, and pre-RT tumor volume) had AUC values of 0.71, 0.81, 0.69, and 0.57, respectively. Based on the obtained AUC values, we proceeded with growth rate γ as a potential model-derived predictor of treatment outcomes and calculated the optimal operating point on the ROC curve using the Jaccard index (Fig. 4B). From this analysis, 0.0042 d−1 (doubling time of 165 d) was identified as the cutoff value to differentiate responders from non-responders, such that patients with estimated tumor growth rates ≤0.0042 d−1 were more likely to be responders, while those with growth rates >0.0042 d−1 were more likely to be non-responders. Although the ability of γ to correctly identify true responders was somewhat low (sensitivity = 61.8%), its ability to correctly identify true non-responders was quite high (specificity = 85.8%), leading to an overall prediction accuracy of 80.5%. The results were validated using the LOOCV method, which led to an average AUC value of 0.816 ± 0.0032 calculated across the 154 ROC curves obtained by retraining the classifier 154 times upon removing a new data point every time (Fig. 4C). The ability of the classifier to correctly predict the left-out data point during LOOCV led to a specificity of 83.3%, sensitivity of 52.9%, and an overall accuracy of 76.6%. The corresponding confusion matrices are shown in Supplementary Fig. 3. Binary classification of mathematical model-based classifier. A ROC curves for α, γ, D50eff, and pre-RT tumor volume used independently as classifiers of local control (LC) outcome. AUC values are shown in the legend. B ROC curve for the γ classifier, highlighting the optimal operating point (red circle) and corresponding performance metrics (inset). C Leave-one-out cross-validation (LOOCV) of the γ classifier applied to the entire cohort (N = 154), with individual ROC curves shown. Confusion matrices for LOOCV results are presented in Supplementary Fig. 3. Survival curves stratified by qEASL growth rate. D, E Kaplan–Meier curves for local control (D) and overall survival (E) stratified by the model-derived tumor growth rate (γ), using the threshold from (B). F Spearman’s correlation between γ and the percentage change in enhancement volume based on qEASL measurements, linking model-predicted dynamics to imaging-derived response assessment. The classifier was further tested for its ability to predict survival using Kaplan-Meier survival curves, log-rank test, and Cox analysis. The survival curves for OS and LC stratified by the growth rate-based binary classifier are shown in Fig. 4D, E. A clear separation was observed between the two cohorts for both OS and LC (P < 0.001 for both). The median OS for responders and non-responders was 36 months (CI 24-48) and 14 months (CI 12-17), respectively. The median LC for non-responders was 19 months (CI 15-29) and the median LC for responders was not reached at last follow-up. In univariate Cox analysis, compared to as a non-responders, classification as a responder was associated with an HR of 0.40 (CI 0.26–0.61; P < 0.001) for death and a HR of 0.29 (CI 0.15–0.56; P < 0.001) for local progression. (Supplementary Table 4). Finally, to establish the validity of the model-based γ classifier to work on the qEASL-based measure of treatment response or as a surrogate for the (%) change in the viable volume classifier, we assessed the correlation between model-derived γ and the percentage change in viable volume. As shown in Fig. 4F, a strong monotonic correlation was observed between the two variables, indicated by Spearman’s rank correlation coefficient ρ = 0.85, and defined by the function yx=0.0153·1−e−0.0047·x+0.01. This suggests that the model-predicted γ values are likely to correctly classify the patient outcomes evaluated using the qEASL-based response measure. Model-based treatment optimization As shown in Fig. 5, the model was tested as a means to optimize or personalize the treatment regimen to improve outcomes. In Fig. 5A, B, the dashed black lines represent the tumor enhancement volume kinetics of the average responders and non-responders examined in Fig. 3C, D. The model was used to simulate different clinically relevant treatment protocols and to predict the corresponding tumor enhancement volume kinetics while keeping the model parameter values constant and equal to the values for the average subjects. As observed, for both responders and non-responders, the only treatment regimen that outperformed the average behavior was 100 Gy in 25 fractions (BED10 140 Gy). Simulations were conducted to assess the differential impact of latency to RT starting from the time of initial radiographic imaging by examining the predicted change in enhancement on day 60 following the conclusion of RT (Supplementary Table 6). Specifically, the model predicted that starting RT seven days sooner may benefit patients who ultimately do not respond to RT. Treatment optimization simulations. Model-based predictions of tumor enhancement volume kinetics corresponding to (A, C) the average responder and (B, D) average non-responder subject under different treatment conditions. Dashed black lines represent the response kinetics corresponding to the average subject simulated in Fig. 4C, D. In (C, D) treatment was initiated seven days sooner than in (A, B). Red arrow indicates the initiation time of treatment, and the other colored arrows denote the termination time of a given treatment (described in legend). Discussion Early identification of responders and non-responders to RT for patients with iCCA has been elusive, owing in part to suboptimal prediction while using traditional size-based radiographic criteria. Without prompt assessment of the response to treatment, clinicians cannot make informed decisions about the most appropriate modifications to a patient’s treatment course. In the current study, we developed a mechanistic model using changes in tumor enhancement following RT to better understand tumor growth kinetics and accurately predict the response to treatment. We found that tumor viability, as estimated by the degree of post-contrast enhancement, outperformed the traditional size-based criteria and effectively stratified patients with respect to LC and OS following RT, independent of other clinical characteristics, treatment parameters, or prognostic features. Furthermore, mathematical modeling of viable tumor volumes revealed earlier initiation of RT as a potential strategy to overcome anticipated non-response to RT. Several prior mechanistic models of radiotherapy response have incorporated detailed biological processes such as repair kinetics, proliferative delays, and biomarker dynamics28–31. These frameworks provide valuable biological insight but generally require dense longitudinal datasets with multiple post-treatment observations in order to fit and validate multi-parameter structures. In contrast, our clinical cohort included only two imaging timepoints per patient (pre- and post-RT), which precluded the direct application of such models without risking identifiability issues. We therefore deliberately adopted a parsimonious approach that estimates only two patient-specific parameters (γ and α), while fixing other terms within biologically justified ranges, to ensure tractability under sparse imaging conditions. Importantly, our model also integrates qEASL-derived perfusion (via BVF) into the potency term D50eff, thereby linking functional imaging biomarkers with mechanistic modeling of treatment response. This design preserves interpretability, captures heterogeneity in outcomes, and extends mechanistic RT modeling to settings where only limited imaging data are available, a context common in routine oncology practice. Personalization of RT dose and fractionation is a captivating potential application of the results of the current study. The developed mathematical model provides a mechanistic description of tumor response to radiation and helps identify a translatable outcome predictor for prospective clinical applications. The model accounts for tumor growth dynamics and radiation-induced tumor death, while considering the role of tumor perfusion (BVF) and other molecular factors (α) in radiation resistance. Since the model is built on qEASL-derived data for viable tumor volume, it may enable more accurate prediction of radiation response, given that the functional or viable volume of the tumor is a more realistic determinant of treatment response, unlike the gross size measured via RECIST, which did not predict outcomes as well as qEASL. Therefore, through characterization of tumor- and treatment-specific parameters from routine clinical measurements or population averages, our model can predict viable tumor volume trajectories, which can provide the basis for treatment design optimization or adaptation in real time for improved outcomes. Importantly, modeling-based analysis revealed the significance of tumor growth rates in determining treatment outcomes, as demonstrated by the high accuracy of γ (AUC of ROC = 0.81) in classifying patient responses. This suggests that a net positive growth rate is maintained by tumors of non-responders, which can perhaps be attributed to molecular factors that upregulate cell proliferation or induce anti-apoptotic genes32. This is consistent with the literature, where pre-treatment tumor growth rates in various solid tumors have been found to negatively affect treatment outcomes with immune checkpoint inhibitors33,34 and chemoradiotherapy35–37. We foresee the application of this finding to estimate tumor growth rate using pre-treatment imaging, thereby predicting treatment outcomes and allowing adaptation of treatment design for improved outcomes. Furthermore, as a representative example, through model simulations, we demonstrated that longer fractionation schedules (25 fractions), combined with higher doses (100 Gy), appeared to be correlated with greater decreases in tumor enhancement, capable of transforming the average non-responder into a responder. These results are consistent with prior work showing a potential LC and OS benefit to ablative doses of RT3. We also show that the timing of RT may have a significant impact on change in enhancement following RT, particularly for the group of patients who were deemed to be non-responders. As a next step, identification of non-responders using pre-RT CT-based features will be crucial in further investigating if a cohort of patients for whom more prompt treatment, longer fractionation, or higher doses may be warranted. As shown in our study population, radiographic assessment using size-based criteria does not provide a complete picture of response to therapy. Consideration of post-treatment morphological changes is essential for a deeper understanding of treatment effects and the different patterns of tumor response. With tumor evaluation using RECIST, most patients in this study (82%) would have been categorized as “stable disease”, despite the wide variability in both tumor control and survival in this cohort. This highlights the relative insensitivity of size-based response criteria in predicting outcomes. Conversely, qEASL has been shown to have higher sensitivity and specificity in estimating tumor response than size-based assessments38. With the present analysis, we have now extended the findings of qEASL response to patients with iCCA treated with RT, supporting the idea that enhancement metrics in hepatobiliary cancers are linked to disease biology and treatment response24,38–42. As an exploratory analysis, we reviewed molecularly characterized patients and identified two potential relationships between mutation status and changes in enhancement. Although the subgroup sample sizes were small, ARID1A mutation was associated with an increase in enhancement volume. Outcome data for patients with ARID1A mutations in biliary tract cancers are limited, but a few studies have suggested a poorer prognosis for these patients, which is consistent with the observed increases in enhancement43,44. Further data are needed to more robustly characterize the associations between molecular profile and treatment response. We acknowledge that our study was limited by its retrospective nature and the number of patients. This cohort spanned 20 years, during which time treatment strategies have evolved. As such, there is heterogeneity in radiation doses and techniques, chemotherapeutic agents, and criteria for patient selection for therapy, which may limit the generalizability of our findings. Interpretation of the molecular profiling data also warrants caution. Although most profiling was performed at diagnosis, in some cases it was obtained later in the disease course, such as at the time of biopsy-confirmed recurrence, introducing variability in timing. This heterogeneity may affect observed associations between mutation status and imaging-based treatment response, particularly in the context of tumor evolution or treatment-driven selection pressures. Furthermore, molecular profiling became more consistently adopted in later years, raising the possibility of temporal confounding. Together, these factors highlight the importance of prospective studies to validate the associations between molecular features and imaging-derived biomarkers of response. Another limitation is the potential for informative censoring due to patient death, which may bias estimates of tumor growth parameters and their association with survival. While we mitigated early censoring by including only patients with at least two imaging time points (pre- and post-treatment), the current model does not explicitly account for the potential dependency between longitudinal tumor dynamics and survival outcomes. Future work incorporating joint modeling frameworks, such as those proposed by Desmée et al.45 and Mbogning et al.46, may help correct for this bias and improve the robustness of prognostic inference. Nevertheless, the ability to measure enhancement over a long period of time despite the use of different treatment modalities is also a strength of the methodology, indicating that assessment by qEASL can be used over a variety of acquisition parameters and clinical practices. Our future directions include independent validation of the reported findings, improvement of tumor registration accuracy through the implementation of biomechanical-registration modeling, ascertainment of the impact of upfront systemic therapy, and investigation of the enhancement pattern mapping of these tumors47–49. Additionally, future work may explore the use of nonlinear mixed-effects (NLME) modeling frameworks to estimate population-level tumor kinetics and quantify inter-individual variability. While our current approach focused on individualized modeling to enable patient-specific prediction and interpretation, NLME methods could complement this by improving parameter stability in sparse data settings and supporting simulation of treatment scenarios across a broader patient spectrum. Finally, while our modeling framework was designed around qEASL-based enhancement volumes due to their ability to capture viable tumor kinetics, future work may explore analogous modeling using volumetric RECIST-derived metrics where appropriate. This would enable comparative evaluation of parameter estimates and outcome associations across different response criteria. In summary, mathematical models utilizing tumor enhancement volume changes following RT may be used to assess response to therapy as early as the first follow-up scan. This measurement outperforms RECIST, which is the standard criterion used for assessing radiographic response. In addition, measurement of tumor growth rate pre-treatment can serve as a predictor of treatment outcomes, which, upon integration with the mathematical model, can be used for treatment personalization. Further validation of these findings using prospective data will help establish viable volume changes as an early signal of treatment response, which may better guide the subsequent management of iCCA patients treated with RT. Methods Patient selection We retrospectively identified 154 patients with biopsy-proven unresectable iCCA treated with definitive radiotherapy (RT) at MD Anderson Cancer Center from 2001–2021. Eligible patients had pre-RT and post-RT liver-protocol computed tomography (CT) scans, including non-contrast and portal venous phases. Overall survival (OS) was measured from the initiation of RT until death or last follow-up. Local control (LC) was defined as the time from the start of RT to local tumor progression, defined as new or progressive tumor within or at the margin of the RT field, or last follow-up. Imaging studies suggestive of local progression were reviewed and confirmed by a radiologist. Notably, RECIST-defined progression and local progression were not treated as equivalent: local progression was defined specifically as radiographic evidence of in-field tumor growth, confirmed by radiologist review, regardless of whether RECIST criteria for progressive disease were met. The study was approved by the Institutional Review Board at MD Anderson Cancer Center (protocol PA14-0646), with a waiver of informed consent. Imaging analysis All patients underwent pre- and post-RT contrast-enhanced liver protocol CT imaging. We performed image visualization, processing, and analysis on the portal venous phase images of both CT scans using Philips IntelliSpace Portal 8 software (Philips Healthcare, Amsterdam, Netherlands; Fig. 1A–D. After image registration, the corresponding non-contrast scan was subtracted from each post-contrast study to remove any artifactual enhancement (e.g., due to calcification, biliary stents, or other metal artifact). The tumor volume was manually contoured on each 5 mm slice with exclusion of visually identifiable arteries or veins to reduce the possibility of contrast within larger vessels confounding measurements of tumor enhancement, which could potentially lead to an artifactually increased estimate of viable tumors. A reference region of non-tumor hepatic parenchyma was selected for each CT study. Viable tumor was classified as any tissue with enhancement exceeding one standard deviation of the mean enhancement measured in Hounsfield units from this reference region. Tumor contours and reference regions were independently verified by a radiation oncologist who was blinded to all patient clinical data and outcomes. To quantify radiographic treatment response, the percent change in viable tumor volume was calculated using qEASL-derived enhancement volumes from the pre-RT baseline and the first post-RT CT scan. This value was treated as a fixed covariate in all classification and survival analyses and was defined as: This continuous variable was used as a predictor in Cox proportional hazards models for OS and LC, and a threshold of ≥33% decrease was used to define binary response classification in receiver operating characteristic (ROC) analyses50. Molecular correlates of disease response Mutation profiling was performed in 94 (61%) patients using previously described techniques, including next-generation sequencing from solid tumor tissue and/or circulating cell-free DNA to screen for single-nucleotide variants, insertions/deletions, copy number gains, and gene fusions5. Associations between mutations and changes in enhancement were evaluated. Mathematical model development and calibration To understand the mechanistic underpinnings of tumor response to RT in iCCA, we adapted our prior work to develop a mathematical model of tumor growth based on qEASL-derived measurements of viable tumor volumes16,51–53. The model (Eq. 1) describes the time-dependent evolution of the qEASL-derived tumor enhancement volume by accounting for exponential tumor growth and radiotherapy-induced tumor death. The model assumes that tumor enhancement volume kinetics is a surrogate for tumor growth kinetics, such that the rate of change of enhancement volume is given by where Venht is the tumor enhancement volume at time t, γ is the first-order tumor growth rate constant, δmax is the maximum tumor death rate constant (achievable under ideal conditions of radiation dose delivery and radiation sensitivity), Dt is the radiation dose absorbed in the tumor at time t, D50eff is the ‘effective’ half-maximal inhibitory dose of RT (i.e., potency of radiation), and Venh0 is the pre-treatment tumor enhancement volume designated as the initial condition. Note that the radiation-induced tumor death rate is proportional to the residual effect of the delivered radiation dose, modeled as a time-dependent decay function Dt. This effect is assumed to be saturable; hence, a Michaelis-Menten term DtD50eff+Dt is used to govern the rate of treatment-induced cell death. We hypothesized that the potency of radiotherapy (or D50eff) depends on tumor perfusion (i.e., tissue oxygenation)16 and non-specific molecular factors (e.g., activation of transcription factors to induce anti-apoptotic genes and upregulate cell proliferation)32. Therefore, the ‘theoretical’ or in vitro D50 of radiotherapy is corrected for by a penalty factor obtained from the patient-specific tumor blood volume fraction (BVF), such that D50eff=D50BVFα. Here, BVF is a dimensionless quantity estimated as the ratio of tumor enhancement volume to total tumor volume (pre-treatment); hence, BVF always assumes a value between 0 and 1. Here, the exponent α is defined as the radiation resistance coefficient that represents the effect of patient-specific, ‘unknown’ molecular factors on radiation resistance. Therefore, a larger value of α leads to a smaller value of the denominator, thereby increasing D50eff (i.e., reducing potency). The biological efficacy of each fraction was modeled as an exponentially decaying function of time after delivery, and the cumulative residual effect at time t was obtained by summing contributions from all prior fractions: Dt=∑iinST1t≥i(t)·D0·e−k·(t−i), where D0 is the administered dose per fraction, k is the decay rate constant representing the duration of radiation-induced biological effects, 1t≥it=1,t≥i0,t 0.05). Baseline categorical characteristics (sex, AJCC stage, ECOG performance status, intrahepatic satellites, concurrent chemotherapy, and radiation technique [proton vs. photon]) were compared between qEASL-defined responders and non-responders using Fisher’s exact test. Welch’s t-test was used to evaluate differences in sample means for model-derived parameter values between responder groups. For classifier training and validation, LC was also operationalized as a binary variable. Patients achieving LC ≥ 24 months were labeled as responders, and those with local failure or censored before 24 months as non-responders. This threshold was based on prior institutional outcomes data and its use in prior RT studies as a clinically meaningful marker of durable local response. These binary labels served as the ground truth for both qEASL-based and model-derived classifiers. Ethics approval and consent to participate The Institutional Review Board Committee at MD Anderson Cancer Center approved the review of the medical records of these patients (PA14-0646). The need for informed patient consent was waived because this was a retrospective review, and no identifiable patient information was included in this report. This study was conducted in accordance with the ethical standards of the Declaration of Helsinki and its amendments. Supplementary information Supplementary Information Acknowledgements We gratefully acknowledge support from Philips Healthcare, the Khalifa Foundation, the Andrew Sabin Family Fellowship, the Center for Radiation Oncology Research, the Sheikh Ahmed Center for Pancreatic Cancer Research, institutional funds from The University of Texas MD Anderson Cancer Center, GE Healthcare, and Cancer Center Support (Core) Grant P30 CA016672 from the National Cancer Institute to MD Anderson. Dr. Eugene Koay was also supported by the NIH (U01CA214263, R01CA221971, P01CA261669, and U01CA230997), Pancreatic Cancer North America, The Warren Y. Soper Charitable Trust, Project Purple, and the Radiological Society of North America (RSD1429). Dr. Brian De was supported by the RSNA Research and Education Foundation (grant number RR2111). Dr. Ethan Ludmir was supported by the Andrew Sabin Family Fellowship. Dr. Prashant Dogra gratefully acknowledges the NIH grant 1R01EB035545. The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript. Author contributions B.D., P.D., M.Z., V.C., E.B.L., and E.J.K. participated in study conception and design. B.D., M.Z., D.E., K.S., A.M.A., and C.W. collected and curated data. B.D., P.D., and M.Z. analyzed and interpreted the data. P.D. performed mathematical modeling and simulations. B.D., P.D., and M.Z. performed statistical analysis. B.D., P.D., M.Z., E.B.L., and E.J.K. drafted the manuscript. All authors critically revised the manuscript for important intellectual content and approved the final manuscript. Data availability Research data are stored in a secure repository at MD Anderson Cancer Center and may be shared upon reasonable request to the corresponding author (E.J.K.), subject to institutional approvals and patient confidentiality considerations. Code availability The custom MATLAB code used for tumor enhancement modeling is publicly available on GitHub at https://github.com/prashantdogra17/Radiotherapy-modeling. Competing interests B.D. reports honoraria from Sermo, Inc. and funding from the Radiological Society of North America. E.B.H. reports grants from Merck Serono. A.C.K. reports ownership of shares in Aravive, Inc. P.D. reports honoraria from ASTRO, ASCO, Imedex, and Bayer. E.J.K. reports grants from National Institutes of Health, Stand Up 2 Cancer, MD Anderson Cancer Center, Elekta, Varian, and GE Healthcare; personal fees from RenovoRx, Nanobiotix, AstraZeneca, and Taylor and Francis; and a consulting/advisory role with Augmenix and Kallisio. All reported conflicts are outside the current work. All other authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. These authors contributed equally: Brian De, Prashant Dogra, Mohamed Zaid. Contributor Information Ethan B. Ludmir, Email: ebludmir@mdanderson.org Eugene J. Koay, Email: EKoay@mdanderson.org Supplementary information The online version contains supplementary material available at 10.1038/s41540-025-00616-9. References Associated Data Supplementary Materials Supplementary Information Data Availability Statement Research data are stored in a secure repository at MD Anderson Cancer Center and may be shared upon reasonable request to the corresponding author (E.J.K.), subject to institutional approvals and patient confidentiality considerations. The custom MATLAB code used for tumor enhancement modeling is publicly available on GitHub at https://github.com/prashantdogra17/Radiotherapy-modeling.