Robust prostate disease classification using transformers with discrete representations.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Machine Learning

Background:

• Convolutional Neural Networks (CNNs) show promise in automated prostate disease classification using multi-parametric MRI.
• Vision Transformers (ViTs), a CNN-free architecture, excel in some image classification tasks but struggle with textural variations common in MRI due to differing acquisition protocols.
• Generalizing MRI models to different magnet strengths (e.g., 1.5 T vs. 3 T) remains a challenge.

Purpose of the Study:

• To develop a more robust vision transformer (ViT) framework for prostate cancer classification on MRI.
• To improve the generalization ability of ViT models to varying magnetic field strengths and acquisition protocols.
• To address the vulnerability of ViTs to textural shifts in MRI data.

Main Methods:

• Proposed a novel framework employing discrete data representations via vector quantization to enhance ViT robustness.
• Input to the transformer model consists of a subset of these discrete representations.
• Utilized cross-attention to integrate discrete representations from T2-weighted and apparent diffusion coefficient (ADC) MRI images.

Main Results:

• The proposed method demonstrated state-of-the-art (SOTA) performance in classifying prostate MRI lesions.
• Achieved superior robustness against domain shifts (e.g., 1.5 T vs. 3 T scanners) compared to existing CNN and transformer models.
• Showcased improved resilience to input space perturbations.

Conclusions:

• Developed an effective method to enhance the robustness of transformer-based prostate lesion classification on MRI.
• The use of discrete representations of T2-weighted and ADC images significantly improves model generalization across different MRI scanner strengths.
• The framework offers a promising solution for reliable automated disease classification in the presence of MRI data variability.