Related Experiment Video
Updated: Jan 14, 2026
A Cognitive Fusion-guided Prostate Biopsy Using Multiparametric Magnetic Resonance Imaging and Transrectal Ultrasound
Published on: March 21, 2025
Adaptive ensemble learning for prostate cancer classification on multi-modal MRI: reducing unnecessary biopsies
Samet Aymaz1, Nur Kara Oğuz2, Şeyma Aymaz3
1Department of Computer Engineering, Trabzon University, Trabzon, Türkiye. sametaymaz@trabzon.edu.tr.
Purpose:
This study aimed to develop and evaluate an adaptive weighted ensemble learning model using multiple CNN feature extractors for multi-modal MRI classification of PI-RADS 3-5 prostate lesions. The primary goal was to reduce unnecessary invasive biopsies while maintaining high diagnostic accuracy for prostate cancer detection.
Methods:
A retrospective diagnostic accuracy study analyzed 196 patients (mean age 64 ± 8.5 years) with PI-RADS 3-5 lesions who underwent multiparametric MRI and biopsy between January 2023-November 2024. Five CNN feature extractors (MobileNet_v2, VGG16, DenseNet121, EfficientNet_b0, ResNet50) were compared within an adaptive weighted ensemble model integrating DCE, DWI, and T2-weighted sequences. The model incorporated expert architectures (CNN, Transformer, Attention LSTM) for each modality with dynamic weighting mechanisms. Performance was evaluated using 5-fold cross-validation with data augmentation and ADASYN balancing, comparing against histopathological reference standards and radiologist interpretations.
Results:
VGG16 achieved the highest diagnostic accuracy (99.0 ± 0.7%, AUC 99.9 ± 0.1%), followed by MobileNet_v2 (97.5 ± 0.7%, AUC 99.7 ± 0.2%). The ensemble model demonstrated superior specificity compared to radiologists' biopsy recommendations for PI-RADS 3-5 lesions (98.9% vs. 0.0%) while maintaining high sensitivity (99.1% vs. 100%). Learned modality weights showed DCE contributed most significantly (41.6 ± 2.0%), followed by T2-weighted (33.9 ± 2.1%) and DWI (24.6 ± 1.6%) sequences.
Conclusion:
The proposed adaptive weighted ensemble model achieved superior diagnostic performance for prostate cancer classification compared to radiologist interpretation, demonstrating significant potential to reduce unnecessary biopsies while maintaining high sensitivity for cancer detection. These findings highlight the potential of the approach to improve the efficiency of prostate cancer diagnosis and support better clinical decision-making in prostate cancer management.
