KEY POINTS
- The retrospective study analyzed longitudinal 0.35-T MR-Linac imaging from 761 prostate cancer patients treated between 2018 and 2025. Most patients (719/761, 95%) received five fractions, generally separated by approximately 2 days.
- Investigators trained a Siamese 3D convolutional neural network with a ResNet-18 architecture to determine which of two fraction images was acquired earlier. Training used 2,238 image pairs from 457 patients, while the held-out test set contained 152 patients and 732 pairs.
- For first-versus-last fraction pairs, the model achieved 95% accuracy (95% CI 93–98%) and AUC 0.99, significantly outperforming an experienced radiologist, who achieved 82% accuracy (p<0.0001).
- When trained across all fraction combinations, performance remained high at 91% accuracy and AUC 0.97. In contrast, performance on simulation-versus-first-fraction images—both obtained before therapeutic radiation exposure—fell to 39% accuracy and AUC 0.34, supporting but not proving that the learned signal was related to treatment.
- The magnitude of AI-detected change increased with time between fractions: model logits correlated with fraction interval at r=0.59 (p<0.0001). Changes varied substantially between individual patients, suggesting heterogeneous longitudinal imaging responses.
- Among 577 patients with available biochemical-recurrence status, those who later developed recurrence showed a significantly flatter trajectory of model logits during treatment. The association remained significant after adjustment for baseline PSA, Grade Group, risk group, T stage and ADT history (p<0.05), but the study did not validate a predictive classifier for recurrence.
- Saliency analysis most often highlighted the bladder lumen (20.5%) and pubic symphysis (19.2%), with frequent prostate involvement. Quantitative analysis also showed that the prostate enlarged and darkened during treatment, while bladder volume and signal intensity decreased; the authors caution that saliency does not establish a biological mechanism.
CLINICAL TAKEAWAY
Routine MR-Linac images may contain substantially more treatment-response information than is apparent to the human eye. The association with biochemical recurrence is particularly interesting, but this remains a single-platform retrospective proof-of-concept; prospective validation is needed before these AI-derived temporal features can guide adaptation or treatment decisions.