PCa-RadHop: A transparent and lightweight feed-forward method for clinically significant prostate cancer segmentation

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Prostate cancer diagnosis relies on PI-RADS, which has a high false positive rate, increasing costs and patient discomfort.
• Deep learning (DL) models offer high segmentation performance but suffer from large size, complexity, and lack of interpretability, being perceived as 'black-boxes'.
• There is a need for more transparent and efficient diagnostic tools in prostate cancer detection.

Purpose of the Study:

• To introduce the PCa-RadHop pipeline for transparent feature extraction in prostate cancer detection.
• To leverage the Green Learning (GL) paradigm for a small model size and low complexity.
• To reduce the false positive rate in prostate cancer diagnosis using bi-parametric Magnetic Resonance Imaging (bp-MRI).

Main Methods:

• The PCa-RadHop pipeline employs a two-stage approach using Green Learning (GL).
• Stage-1 extracts radiomics features from bp-MRI to predict an initial heatmap.
• Stage-2 refines predictions by incorporating contextual information and radiomics features from detected Regions of Interest (ROIs) to minimize false positives.

Main Results:

• The PCa-RadHop pipeline demonstrated competitive performance against deep learning models on the PI-CAI dataset.
• Achieved an area under the curve (AUC) of 0.807 in a cohort of 1,000 patients.
• Maintained significantly smaller model size and complexity compared to existing DL models.

Conclusions:

• PCa-RadHop offers a transparent and efficient alternative for prostate cancer diagnosis.
• The method effectively reduces false positives while maintaining high diagnostic performance.
• The pipeline's small model size and low complexity make it a practical tool for clinical application.