Incorporating functional soft tissue deformations in AI model training for spatially accurate prostate cancer...

Area of Science:

• Medical Imaging and Artificial Intelligence
• Prostate Cancer Diagnostics

Background:

• Artificial intelligence (AI) systems for prostate cancer detection require high performance and generalization.
• Prostate imaging is affected by anatomical variations due to physiological changes in surrounding organs like the bladder and rectum.

Purpose of the Study:

• To enhance AI prostate cancer detection systems by simulating physiological size changes of the bladder and rectum.
• To improve the performance and generalization ability of AI models by accounting for associated prostate and lesion deformations.

Main Methods:

• A retrospective study of 1028 bi-parametric MRI examinations was conducted.
• An 'anatomy-informed' transformation was integrated into the nnU-Net training, simulating prostate deformations from bladder and rectal size changes.
• Performance was evaluated using free-response receiver operating characteristic (FROC), weighted alternative FROC (wAFROC), and receiver operating characteristic (ROC/LROC) analyses.

Main Results:

• The anatomy-informed model significantly increased lesion-level detection of true positive lesions by 18.8% on the independent test set (p=0.01).
• Significantly higher performance was observed in wAFROC analysis (0.597 to 0.639, p<0.01).
• Localized ROC (LROC) analysis demonstrated increased performance (0.471 to 0.546).

Conclusions:

• Simulating bladder and rectal size variations during AI model training significantly improved lesion detection performance.
• This approach is crucial for accurate lesion localization, supporting diagnostics and therapeutic interventions like MRI-guided biopsies or focal therapy.