Adaptive ensemble learning for prostate cancer classification on multi-modal MRI: reducing unnecessary biopsies

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Prostate cancer diagnosis relies on multiparametric MRI (mpMRI) and biopsy.
• Accurate classification of PI-RADS 3-5 lesions is crucial to avoid unnecessary invasive procedures.
• Current diagnostic methods can be improved for better patient outcomes.

Purpose of the Study:

• To develop and evaluate an adaptive weighted ensemble learning model for multi-modal MRI classification of PI-RADS 3-5 prostate lesions.
• To enhance prostate cancer detection accuracy and reduce invasive biopsies.
• To integrate multiple Convolutional Neural Network (CNN) feature extractors for improved diagnostic performance.

Main Methods:

• A retrospective study analyzed 196 patients with PI-RADS 3-5 lesions.
• Five CNN feature extractors (MobileNet_v2, VGG16, DenseNet121, EfficientNet_b0, ResNet50) were employed within an adaptive weighted ensemble model.
• The model integrated dynamic weighting of DCE, DWI, and T2-weighted MRI sequences, evaluated using 5-fold cross-validation with data augmentation and ADASYN balancing.

Main Results:

• VGG16 achieved the highest diagnostic accuracy (99.0%) and AUC (99.9%).
• The ensemble model showed superior specificity (98.9%) compared to radiologists' recommendations, maintaining high sensitivity (99.1%).
• Dynamic weighting highlighted DCE (41.6%), T2-weighted (33.9%), and DWI (24.6%) sequences as most significant.

Conclusions:

• The adaptive weighted ensemble model demonstrated superior diagnostic performance for prostate cancer classification.
• The model shows significant potential to reduce unnecessary prostate biopsies while maintaining high sensitivity for cancer detection.
• This approach can improve prostate cancer diagnosis efficiency and support clinical decision-making.