Deep learning-based artificial intelligence for prostate cancer detection at biparametric MRI

Area of Science:

• Radiology and Medical Imaging
• Artificial Intelligence in Medicine
• Oncology

Background:

• Prostate cancer detection relies heavily on MRI interpretation.
• Accurate identification of clinically significant prostate cancer (≥ISUP1) is crucial.
• Developing automated systems can enhance diagnostic efficiency and accuracy.

Purpose of the Study:

• To develop and evaluate a fully automated deep learning (DL) system for prostate cancer lesion detection and segmentation using MRI.
• To assess the performance of DL models (UNet and AH-Net) in identifying prostate cancer lesions with histopathological confirmation.

Main Methods:

• Utilized MRI scans from two institutions for training, validation, and testing.
• Contoured MRI-visible lesions and used histopathology results as ground truth.
• Trained UNet and AH-Net architectures for lesion detection and segmentation, aiming to detect cancer ≥ISUP1.

Main Results:

• The AH-Net model demonstrated higher Dice coefficients (0.403 training, 0.307 validation) compared to UNet (0.372 training, 0.287 validation).
• In the validation set, AH-Net achieved 74.4% sensitivity and 47.8% PPV, outperforming UNet (70.9% sensitivity, 35.5% PPV).
• In the test set, UNet had higher sensitivity (72.8%) than AH-Net (63.0%), with AH-Net showing fewer false positives per patient (1.40 vs 1.90).

Conclusions:

• A DL-based AI approach for prostate cancer detection on biparametric MRI demonstrates reasonable performance.
• False positive lesion identification remains a challenge for AI-assisted detection algorithms.
• The developed system can serve as a valuable adjunct tool for radiologists in prostate cancer diagnosis.