Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Cross-Slice Attention and Evidential Critical Loss for Uncertainty-Aware Prostate Cancer Detection

Alex Ling Yu Hung, Haoxin Zheng, Kai Zhao, Kaifeng Pang, Demetri Terzopoulos, Kyunghyun Sung

TL;DR

This work tackles uncertainty-aware prostate cancer detection from mpMRI by combining a novel 2.5D cross-slice attention mechanism, Global-Local Cross-Slice Attention (GLCSA), with an evidential critical loss (EC) to enable Evidential Deep Learning (EDL) for per-pixel uncertainty estimation. GLCSA integrates semantic, positional, and slice-wise attention to fuse global and local information across slices, and is designed to slot into UNet-like architectures at skip connections. The EC loss focuses learning on hard lesion pixels by modulating the standard evidential objective and adding a KL regularizer, enabling robust detection of small cancer lesions under highly imbalanced data. On two datasets, the method achieves state-of-the-art Free-Response Operating Characteristic (FROC) performance and provides superior epistemic uncertainty calibration compared with MC dropout, suggesting improved reliability and potential clinical utility in guiding biopsies and radiologist workflows.

Abstract

Current deep learning-based models typically analyze medical images in either 2D or 3D albeit disregarding volumetric information or suffering sub-optimal performance due to the anisotropic resolution of MR data. Furthermore, providing an accurate uncertainty estimation is beneficial to clinicians, as it indicates how confident a model is about its prediction. We propose a novel 2.5D cross-slice attention model that utilizes both global and local information, along with an evidential critical loss, to perform evidential deep learning for the detection in MR images of prostate cancer, one of the most common cancers and a leading cause of cancer-related death in men. We perform extensive experiments with our model on two different datasets and achieve state-of-the-art performance in prostate cancer detection along with improved epistemic uncertainty estimation. The implementation of the model is available at https://github.com/aL3x-O-o-Hung/GLCSA_ECLoss.

Cross-Slice Attention and Evidential Critical Loss for Uncertainty-Aware Prostate Cancer Detection

TL;DR

This work tackles uncertainty-aware prostate cancer detection from mpMRI by combining a novel 2.5D cross-slice attention mechanism, Global-Local Cross-Slice Attention (GLCSA), with an evidential critical loss (EC) to enable Evidential Deep Learning (EDL) for per-pixel uncertainty estimation. GLCSA integrates semantic, positional, and slice-wise attention to fuse global and local information across slices, and is designed to slot into UNet-like architectures at skip connections. The EC loss focuses learning on hard lesion pixels by modulating the standard evidential objective and adding a KL regularizer, enabling robust detection of small cancer lesions under highly imbalanced data. On two datasets, the method achieves state-of-the-art Free-Response Operating Characteristic (FROC) performance and provides superior epistemic uncertainty calibration compared with MC dropout, suggesting improved reliability and potential clinical utility in guiding biopsies and radiologist workflows.

Abstract

Current deep learning-based models typically analyze medical images in either 2D or 3D albeit disregarding volumetric information or suffering sub-optimal performance due to the anisotropic resolution of MR data. Furthermore, providing an accurate uncertainty estimation is beneficial to clinicians, as it indicates how confident a model is about its prediction. We propose a novel 2.5D cross-slice attention model that utilizes both global and local information, along with an evidential critical loss, to perform evidential deep learning for the detection in MR images of prostate cancer, one of the most common cancers and a leading cause of cancer-related death in men. We perform extensive experiments with our model on two different datasets and achieve state-of-the-art performance in prostate cancer detection along with improved epistemic uncertainty estimation. The implementation of the model is available at https://github.com/aL3x-O-o-Hung/GLCSA_ECLoss.
Paper Structure (17 sections, 7 equations, 3 figures, 2 tables)

This paper contains 17 sections, 7 equations, 3 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Preliminaries
  3. 2.5D Volumetric Image Analysis
  4. Evidential Deep Learning
  5. Global-Local Cross-Slice Attention
  6. Semantic Attention
  7. Positional Attention
  8. Slice Attention
  9. Evidential Critical Loss
  10. Experiments
  11. Datasets
  12. Implementation Details
  13. Quantitative Results
  14. Ablation Study
  15. Uncertainty Estimation
  16. ...and 2 more sections

Figures (3)

  • Figure 1: GLCSA implementation.
  • Figure 2: Quantitative results.
  • Figure 3: (a) Comparison of EDL and MC dropout on uncertainty calibration, and (b) FROC curves for different uncertainty thresholds.