Prostate cancer classification using 3D deep learning and ultrasound video clips: a multicenter study

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Oncology

Background:

• Prostate cancer diagnosis relies on various imaging modalities.
• Transrectal ultrasound (TRUS) provides real-time imaging during prostate examinations.
• Developing accurate AI tools for TRUS analysis can improve diagnostic efficiency.

Purpose of the Study:

• To evaluate the effectiveness of deep learning models in predicting prostate cancer using TRUS video clips.
• To compare the performance of a deep learning model against other AI models and human expert diagnoses.

Main Methods:

• Manual segmentation of TRUS video clips from 815 men.
• Development and internal validation of a deep learning-inflated 3D ConvNet (I3D) model.
• External validation on two independent test sets and comparison with ResNet 50, ML models, and sonologists.

Main Results:

• The I3D model achieved diagnostic classification AUCs > 0.86 across internal and external test sets.
• I3D demonstrated superior consistency in sensitivity, specificity, and accuracy (kappa > 0.62).
• The I3D model significantly outperformed other AI models and sonologists in prostate cancer prediction (p<0.05).

Conclusions:

• Deep learning models, specifically the I3D model, are valuable tools for classifying and predicting prostate cancer using TRUS video clips.
• The I3D model shows potential for enhancing diagnostic accuracy and consistency in prostate cancer detection.