Prostate cancer detection from multi-institution multiparametric MRIs using deep convolutional neural networks

Area of Science:

• Radiology
• Artificial Intelligence
• Oncology

Background:

• Multiparametric magnetic resonance imaging (mpMRI) is crucial for early prostate cancer diagnosis but faces interpretation challenges and interreader variability.
• Computer-aided diagnostic (CAD) systems can assist radiologists in identifying suspected cancerous lesions on mpMRI.

Purpose of the Study:

• To develop and optimize a deep convolutional neural network (CNN) for generating probability maps to aid prostate cancer detection on mpMRI.
• To evaluate the performance of the optimized CNN in identifying prostate cancer lesions.

Main Methods:

• Optimized a holistically nested edge detection (HED) deep CNN using T2, apparent diffusion coefficient, and high b-value MRI images from 186 patients.
• Utilized expert-drawn tumor segmentations based on histologic evidence as ground truth.
• Evaluated slice-level probability maps at the lesion level, comparing performance between peripheral and transition zones.

Main Results:

• The optimized HED model, using 5x5 convolutional kernels and Adam optimizer, achieved a high Area Under the Curve (AUC) of 0.94 ± 0.01 in the peripheral zone.
• The CAD system demonstrated significantly better performance (p < 0.001) in the peripheral zone compared to the transition zone.
• The developed CAD system outperformed a previous version in head-to-head comparisons on endorectal coil (ERC) test cases (AUC = 0.97 ± 0.01).

Conclusions:

• The developed computer-aided diagnostic system significantly improves the prediction of prostate cancer lesions on mpMRI.
• This CAD system establishes a state-of-the-art performance, offering valuable assistance to radiologists in prostate cancer diagnosis.