3D-Vision-Transformer Stacking Ensemble for Assessing Prostate Cancer Aggressiveness from T2w Images.

Area of Science:

• Computer Vision
• Medical Imaging
• Artificial Intelligence

Background:

• Vision transformers are advanced AI models typically used for 2D image analysis.
• Current methods for assessing prostate cancer aggressiveness often require invasive biopsies.
• There is a need for non-invasive methods to accurately diagnose cancer severity.

Purpose of the Study:

• To develop and evaluate a trained-from-scratch 3D vision transformer stacking ensemble for assessing prostate cancer aggressiveness.
• To utilize T2-weighted MRI images for non-invasive cancer diagnosis.
• To improve diagnostic accuracy and reduce reliance on biopsies.

Main Methods:

• Trained 18 individual 3D vision transformers (ViTs) on T2-weighted MRI data.
• Combined ViTs into two- and three-model stacking ensembles.
• Employed five-fold cross-validation to evaluate accuracy, prediction confidence, and calibration.
• Optimized base ViTs and compared ensemble performance against base models.

Main Results:

• The best 3D vision transformer stacking ensemble achieved state-of-the-art Area Under the Receiving Operating Curve (0.89).
• The ensemble significantly improved the Area Under the Precision-Recall Curve by 22% compared to the base model (p < 0.001).
• The ensemble demonstrated lower confidence in classifying positive cases despite high accuracy.

Conclusions:

• A 3D vision transformer stacking ensemble can effectively assess prostate cancer aggressiveness from T2-weighted MRI.
• This approach offers a promising non-invasive alternative to traditional biopsy methods.
• Further research is needed to enhance the confidence metrics of the ensemble model for clinical application.