AI-Based Detection of In-Treatment Changes from Prostate MR-Linac Images

Purpose: To investigate whether routinely acquired longitudinal MR-Linac images can be leveraged to characterize treatment-induced changes during radiotherapy, particularly subtle inter-fraction changes over short intervals (average of 2 days). Materials and Methods: This retrospective study included a series of 0.35T MR-Linac images from 761 patients. An artificial intelligence (deep learning) model was used to characterize treatment-induced changes by predicting the temporal order of paired images. The model was first trained with the images from the first and the last fractions (F1-FL), then with all pairs (All-pairs). Model performance was assessed using quantitative metrics (accuracy and AUC), compared to a radiologist’s performance, and qualitative analyses - the saliency map evaluation to investigate affected anatomical regions. Input ablation experiments were performed to identify the anatomical regions altered by radiotherapy. The radiologist conducted an additional task on partial images reconstructed by saliency map regions, reporting observations as well. Quantitative image analysis was conducted to investigate the results from the model and the radiologist. Results: The F1-FL model yielded near-perfect performance (AUC of 0.99), significantly outperforming the radiologist. The All-pairs model yielded an AUC of 0.97. This performance reflects therapy-induced changes, supported by the performance correlation to fraction intervals, ablation tests and expert’s interpretation. Primary regions driving the predictions were prostate, bladder, and pubic symphysis. Conclusion: The model accurately predicts temporal order of MR-Linac fractions and detects radiation-induced changes over one or a few days, including prostate and adjacent organ alterations confirmed by experts. This underscores MR-Linac’s potential for advanced image analysis beyond image guidance.

💡 Research Summary

This retrospective study investigates whether routinely acquired longitudinal MR‑Linac images can be leveraged to detect subtle, treatment‑induced changes during prostate radiotherapy. The authors assembled a cohort of 761 prostate cancer patients treated on a 0.35 T ViewRay MR‑Idian system between 2018 and 2025. Each patient received five treatment fractions (F1‑F5) with an average inter‑fraction interval of two days. Images were acquired using a True Fast Imaging with Steady‑State Free Precession (TRUFI) sequence, cropped to an 80 × 80 × 80 voxel volume centered on a prostate mask, and split patient‑wise into training (60 %), validation (20 %), and test (20 %) sets.

The core algorithm builds on the LILAC framework, employing a Siamese 3D convolutional neural network with a ResNet‑18 backbone. Paired volumes are processed independently, their feature vectors subtracted, and the difference passed through a fully‑connected layer to produce logits for a binary classification task: is the pair presented in the correct temporal order? Training proceeded in two stages. First, a curriculum learning phase used only the first‑and‑last fraction pairs (F1‑FL), which exhibit the largest radiation‑induced differences. Second, the model was fine‑tuned on all possible ordered pairs (All‑pairs) using the weights from the best F1‑FL model.

Performance was evaluated on the held‑out test set (152 patients, 732 pairs) using accuracy and area under the ROC curve (AUC) with 1 000 bootstrap resamples for confidence intervals. The F1‑FL model achieved an accuracy of 0.95 (95 % CI 0.93‑0.98) and an AUC of 0.99 (0.99‑1.00), markedly outperforming a senior radiologist who obtained 0.82 accuracy (0.76‑0.88). The All‑pairs model retained high performance (accuracy 0.91, AUC 0.97). In contrast, when the same model was applied to pre‑treatment simulation‑to‑F1 pairs, accuracy dropped to 0.40 and AUC to 0.34, indicating that the network is specifically sensitive to radiation‑driven changes rather than generic temporal drift.

To probe what anatomical structures drive the predictions, the authors generated Grad‑CAM‑based saliency maps. Consistently across patients, the prostate, bladder, and the pubic symphysis (pelvic bone) emerged as the most salient regions. Input‑ablation experiments further clarified these findings: masking (occluding) the prostate or bladder reduced model performance substantially, while preserving only these organs (only‑organ input) retained most of the predictive power. Mask‑only inputs (binary organ shapes without intensity) yielded intermediate performance, suggesting that both shape and intensity changes contribute.

Quantitative image analysis corroborated the deep‑learning results. Between F1 and FL, the prostate volume decreased modestly (≈5 % median reduction, p < 0.01) and bladder intensity metrics (mean and standard deviation) shifted significantly, reflecting edema or filling variations induced by radiation. Patients receiving androgen deprivation therapy showed no difference in model logits, reinforcing that the network captures radiation effects rather than hormonal changes.

The radiologist was also asked to perform the temporal‑ordering task on images restricted to the saliency‑highlighted regions (saliency‑restricted MR). Their accuracy remained lower than the AI, and their rationale often cited subtle shape changes that were difficult to discern without full context, underscoring the advantage of data‑driven feature extraction.

Overall, the study demonstrates that even low‑field (0.35 T) MR‑Linac images contain sufficient information for a deep‑learning model to detect and quantify treatment‑induced morphological changes over intervals as short as one to two days. The ability to automatically order images implies that the model has learned a monotonic representation of cumulative radiation dose effects. This opens the door to using MR‑Linac data not only for image‑guided positioning but also for real‑time treatment monitoring, adaptive replanning, and potentially early toxicity prediction. Future work should explore multimodal integration (e.g., diffusion‑weighted ADC maps), longitudinal modeling of dose‑response curves, and prospective validation within clinical workflows to translate these AI‑derived biomarkers into actionable decision support tools.