Abstract
Detecting prostate cancer (PCa) in multi-parametric MRI (mpMRI) is a complex task hindered by scan anisotropy, extreme class imbalance, and the need for reliable uncertainty quantification. This paper introduces a novel "2.5D" deep learning framework that integrates a Global-Local Cross-Slice Attention (GLCSA) mechanism with an Evidential Critical (EC) Loss. The GLCSA module captures volumetric context by sequentially applying semantic, positional, and slice-level attention, while the EC Loss utilizes Evidential Deep Learning to provide robust uncertainty scores and focus learning on critical lesion pixels. Evaluated on internal and public PI-CAI datasets, the model achieves state-of-the-art detection performance and superior uncertainty calibration compared to standard 3D models and Monte Carlo Dropout methods.
Introduction
Prostate cancer remains a leading cause of mortality among men, and mpMRI is the primary tool for non-invasive diagnosis. However, AI-driven detection faces three significant hurdles. First, MRI scans are typically anisotropic, with high in-plane resolution but low through-plane resolution, making standard 3D convolutions inefficient. Second, cancer lesions are often minute, leading to a severe class imbalance where background pixels can outnumber lesion pixels by a factor of 10,000. Finally, clinical adoption requires models that can quantify their own uncertainty to avoid overconfident misdiagnoses.
Methodology
Global-Local Cross-Slice Attention (GLCSA)
The researchers propose a "2.5D" approach that processes 2D slices while leveraging volumetric information through the GLCSA mechanism. This module is integrated into the skip connections of a standard nnU-Net architecture, allowing the model to refine features across different scales.
The GLCSA mechanism operates in three distinct stages:
- Semantic Attention: This stage re-weights feature channels based on their global and local importance across the volume, ensuring the model focuses on the most informative semantic features.
- Positional Attention: This identifies relevant spatial regions within each slice, helping the model pinpoint potential lesion locations.
- Slice Attention: This learns to weight individual slices within a stack, prioritizing central slices where the prostate is most likely to be visible over peripheral slices.
Evidential Critical (EC) Loss
To address uncertainty and class imbalance, the framework employs Evidential Deep Learning (EDL). Instead of a standard softmax output, the model predicts parameters for a Dirichlet distribution, allowing it to model "epistemic uncertainty."
The authors further developed the Evidential Critical (EC) Loss, which incorporates a modulation factor similar to Focal Loss. This factor down-weights the loss for "easy" background pixels and amplifies it for "critical" lesion pixels, forcing the model to prioritize the detection of small, hard-to-find tumors.
Experimental Results
The model was validated on an internal dataset of 652 patients and the public PI-CAI benchmark of 1,293 patients.
- Detection Performance: The full framework (nnU-Net + GLCSA + EC Loss) achieved a sensitivity of 0.870 at 3 false positives per patient on the internal dataset, significantly outperforming the baseline nnU-Net (0.801). It also surpassed other state-of-the-art architectures like Swin-UNet and VNet.
- Uncertainty Calibration: The EDL-based approach demonstrated a much stronger correlation between uncertainty and error than Monte Carlo (MC) Dropout. By filtering predictions based on an uncertainty threshold, the researchers were able to further improve the system's sensitivity at low false-positive rates, providing a practical tool for clinical decision support.
Conclusion
By combining a 2.5D cross-slice attention mechanism with evidential learning, this framework effectively addresses the unique challenges of prostate MRI analysis. The GLCSA module provides the benefits of 3D context without the resolution-related pitfalls of 3D convolutions, while the EC Loss ensures the model remains focused on critical lesions and provides reliable confidence scores. Future work will explore the integration of these uncertainty maps into real-world radiologist workflows to enhance diagnostic accuracy.
Summary updated