AI Model Outperforms Radiologist in Detecting Prostate Radiotherapy Changes on MR-Linac Images
A deep learning model achieved superior performance compared to a radiologist in identifying subtle temporal changes in prostate MR-Linac images during radiotherapy. The retrospective study analyzed longitudinal 0.35T MR-Linac images from 761 patients, employing a temporal ordering method via pairwise comparison. Using first-to-last fraction pairs (F1-FL), the AI model demonstrated exceptional accuracy of 0.95 and an AUC of 0.99. Performance assessment included quantitative metrics and qualitative evaluation through saliency maps that highlight anatomical regions associated with imaging changes. The research explored the broader potential of MR-Linac imaging for detecting inter-fraction changes during routine prostate radiotherapy. The model was also trained using all fraction pairs (All-pairs) configuration. This investigation represents significant progress in medical imaging analysis for cancer treatment monitoring.
Key facts
- AI-based method detects subtle inter-fraction changes in MR-Linac images
- Retrospective study included 761 patients
- Used longitudinal 0.35T MR-Linac images
- Deep learning model employed temporal ordering via pairwise comparison
- F1-FL model achieved AUC=0.99 and accuracy=0.95
- Model outperformed radiologist in temporal ordering
- Qualitative evaluation used saliency maps
- Research explores broader potential of MR-Linac imaging
Entities
—