Clinically significant prostate cancer detection with deep learning in a multi-center magnetic resonance imaging...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Early detection of clinically significant prostate cancer (csPCa) is vital for patient outcomes.
• Traditional methods like Digital Rectal Exam and Prostate-Specific Antigen (PSA) tests have limitations in sensitivity and specificity.
• Advanced diagnostic tools are needed for accurate csPCa classification.

Purpose of the Study:

• To develop and validate an AI-based approach for csPCa classification using MRI data.
• To integrate and leverage diverse datasets, including the PI-CAI Challenge and the BIMCV Prostate dataset.
• To enhance the robustness and interpretability of AI models in prostate cancer diagnosis.

Main Methods:

• Utilized a custom-trained nnUNet for prostate segmentation and a 3D EfficientNet-B7 for classification.
• Employed transfer learning by fine-tuning PI-CAI pretrained models on the BIMCV dataset, creating an ensemble meta-learner.
• Implemented data augmentation via synthesized ADC maps and employed interpretability techniques (occlusion sensitivity, guided backpropagation).

Main Results:

• The AI ensemble model achieved an Area Under the Curve (AUC) of 0.816 on an independent hold-out set.
• Significantly outperformed a non-pretrained baseline model (AUC 0.71).
• Demonstrated effective data augmentation using synthesized ADC maps without domain shift and provided model interpretability.

Conclusions:

• AI-enhanced MRI techniques show significant potential for advancing csPCa detection and diagnosis.
• The developed ensemble model offers a robust and accurate method for classifying clinically significant prostate cancer.
• The study highlights the value of diverse datasets and transfer learning in building effective AI diagnostic tools.