Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Deep Learning with Uncertainty Quantification for Predicting the Segmentation Dice Coefficient of Prostate Cancer Biopsy Images

Audrey Xie, Elhoucine Elfatimi, Sambuddha Ghosal, Pratik Shah

TL;DR

This work addresses the need for uncertainty-aware evaluation of deep learning-based histopathology segmentation in clinical settings by linking model uncertainty to segmentation performance. It uses Monte Carlo dropout and backpropagation to generate pixel-level uncertainties for prostate H&E biopsy images and introduces a region-specific uncertainty algorithm to compute ROI-based uncertainties. The study then trains linear models to predict Dice scores from ROI uncertainties (tumor, non-tumor, and non-tissue regions), demonstrating that ROI-based uncertainty provides accurate Dice predictions with low RMSE and stronger correlations than image-wide uncertainty. Results on the Gleason MICCAI 2019 dataset show Dice around $0.899$ and AUROC around $0.894$, with uncertainty maps highlighting boundaries where predictions are less reliable, thereby enabling more informed clinical decision-making. Overall, the approach advances reliable deployment of DLMs in pathology by offering region-aware performance estimation and practical uncertainty visualization.

Abstract

Deep learning models (DLMs) can achieve state-of-the-art performance in histopathology image segmentation and classification, but have limited deployment potential in real-world clinical settings. Uncertainty estimates of DLMs can increase trust by identifying predictions and images that need further review. Dice scores and coefficients (Dice) are benchmarks for evaluation of image segmentation performance, but are usually not evaluated with DLM uncertainty quantification. This study reports DLMs trained with uncertainty estimations, using randomly initialized weights and Monte Carlo dropout, to segment tumors from microscopic Hematoxylin and Eosin dye stained prostate core biopsy histology RGB images. Image-level maps showed significant correlation (Spearman's rank, p < 0.05) between overall and specific prostate tissue image sub-region uncertainties with model performance estimations by Dice. This study reports that linear models, which can predict Dice segmentation scores from multiple clinical sub-region-based uncertainties of prostate cancer, can serve as a more comprehensive performance evaluation metric without loss in predictive capability of DLMs, with a low root mean square error.

Deep Learning with Uncertainty Quantification for Predicting the Segmentation Dice Coefficient of Prostate Cancer Biopsy Images

TL;DR

This work addresses the need for uncertainty-aware evaluation of deep learning-based histopathology segmentation in clinical settings by linking model uncertainty to segmentation performance. It uses Monte Carlo dropout and backpropagation to generate pixel-level uncertainties for prostate H&E biopsy images and introduces a region-specific uncertainty algorithm to compute ROI-based uncertainties. The study then trains linear models to predict Dice scores from ROI uncertainties (tumor, non-tumor, and non-tissue regions), demonstrating that ROI-based uncertainty provides accurate Dice predictions with low RMSE and stronger correlations than image-wide uncertainty. Results on the Gleason MICCAI 2019 dataset show Dice around and AUROC around , with uncertainty maps highlighting boundaries where predictions are less reliable, thereby enabling more informed clinical decision-making. Overall, the approach advances reliable deployment of DLMs in pathology by offering region-aware performance estimation and practical uncertainty visualization.

Abstract

Deep learning models (DLMs) can achieve state-of-the-art performance in histopathology image segmentation and classification, but have limited deployment potential in real-world clinical settings. Uncertainty estimates of DLMs can increase trust by identifying predictions and images that need further review. Dice scores and coefficients (Dice) are benchmarks for evaluation of image segmentation performance, but are usually not evaluated with DLM uncertainty quantification. This study reports DLMs trained with uncertainty estimations, using randomly initialized weights and Monte Carlo dropout, to segment tumors from microscopic Hematoxylin and Eosin dye stained prostate core biopsy histology RGB images. Image-level maps showed significant correlation (Spearman's rank, p < 0.05) between overall and specific prostate tissue image sub-region uncertainties with model performance estimations by Dice. This study reports that linear models, which can predict Dice segmentation scores from multiple clinical sub-region-based uncertainties of prostate cancer, can serve as a more comprehensive performance evaluation metric without loss in predictive capability of DLMs, with a low root mean square error.

Paper Structure

This paper contains 16 sections, 11 equations, 3 figures, 1 table, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. Summary of contributions
  4. Data description
  5. Methods
  6. Model training
  7. Model prediction and evaluation
  8. Results and discussion
  9. Model Performance and Uncertainty Analysis
  10. Qualitative evaluation
  11. Performance Modeling Through Uncertainty Quantification
  12. Tumor-containing images
  13. Tumor-free images
  14. Discussion
  15. Conclusion and Limitations
  16. ...and 1 more sections

Figures (3)

  • Figure 1: Integrated analysis flow: Sequential relationships in histopathology image analysis with deep learning, uncertainty maps, Dice score, and Monte Carlo Dropout (MCD).
  • Figure 2: Segmentation and Uncertainty Maps for Backprop and Monte-Carlo Dropout Augmented Trained-from-Scratch (TFS) models. “GT” – clinical ground-truth, “back-gr” – non-tissue regions.
  • Figure 3: Visualization of model output segmentation and region-based uncertainty maps for prostate tumor segmentation using a Monte Carlo dropout (MCD) deep learning model trained on Hematoxylin and Eosin (H&E) stained prostate biopsy images. Columns left to right:A- Input RGB image; B- Clinical ground-truth binary mask; C- MCD output segmentation binary mask; D- MCD model uncertainty estimate maps; E- Tumor tissue uncertainty map; F- Non-tumor tissue uncertainty map; G- Non-tissue uncertainty map. Color bar shows model uncertainty, with red for higher and blue for lower uncertainty (minimum value 0). "GT" -- clinical ground-truth, "back-gr" -- non-tissue regions.