AI-based detection of temporal changes in MR-Linac images acquired during routine prostate radiotherapy

Abstract

Background and purpose:

To investigate whether an AI-based method can detect subtle inter-fraction changes in MR-Linac images acquired during radiotherapy and explore the broader potential of MR-Linac imaging.

Material and methods:

This retrospective study included longitudinal 0.35T MR-Linac images from 761 patients. To identify temporal changes, we employed a deep learning model using temporal ordering via pairwise comparison, previously shown effective for longitudinal imaging studies. The model was trained using first-to-last fraction pairs ($F_1$-$F_L$) and all pairs (All-pairs). Performance was assessed using quantitative metrics (accuracy and AUC) and compared against a radiologist’s performance. Qualitative evaluation was performed using saliency maps, which identify anatomical regions associated with temporal imaging changes.

Results:

The $F_1$–$F_L$ model demonstrated high performance (AUC=0.99; accuracy=0.95) and outperformed the radiologist in temporal ordering task. The All-pairs model also showed high performance (AUC = 0.97; accuracy = 0.91). The performance was correlated to fractional intervals and was reduced for non-radiation-exposed timepoints ($Sim$ and $F_1$). Patients who later developed biochemical recurrence exhibited smaller longitudinal changes in model logits during radiotherapy. Regions contributing to predictions included the prostate, bladder, and pubic symphysis.

Conclusions:

These findings demonstrate the feasibility of detecting subtle inter-fractional changes over short periods (two days on average) in routine MR-Linac images during prostate radiotherapy and contribute to the development of imaging biomarkers for treatment response.