Uncertainty-Aware Ensemble Segmentation of Breast Cancer Tissue Microarrays

Lucia Schmidt-Santiago, Roman Kinakh, Sergio Carreras-Salinas, Sara Guerrero-Aspizcua, Gonzalo R. Ríos-Muñoz, Arrate Munoz-Barrutia

Keywords: tissue microarrays, deep learning, segmentation, ensemble, uncertainty, histopathology

TL;DR: We propose a robust deep learning pipeline combining ensemble U-Nets with uncertainty estimation to accurately segment breast cancer TMAs, outperforming prior methods and aiding interpretability.

Abstract:

Breast cancer Tissue Microarrays (TMAs) offer a high-throughput platform for studying tumor morphology and biomarker expression. We present an automated deep learning pipeline for semantic segmentation of Hematoxylin and Eosin (H\&E)-stained breast cancer TMAs, integrating ensemble U-Net architectures with ResNet encoders and Monte Carlo Dropout (MCDO) for uncertainty estimation. A robust pre-processing workflow addresses illumination artifacts, staining variability, and tissue detection.

Multiple U-Net models were trained using distinct loss functions to address class imbalance and feature diversity. Predictions were combined via soft voting, emulating consensus among pathologists. Uncertainty was quantified using MCDO across ensemble outputs, enhancing reliability and interpretability.

Our pipeline outperforms similar methods such as WeGleNet ($\text{mIoU}=0.4368$) and HistoSegNet ($\text{mIoU}=0.5505$), achieving a mean IoU of $0.58 \pm 0.11$ and Dice Score of $0.66 \pm 0.10$. Calibration analysis shows superior alignment of standard deviation–based uncertainty estimates with actual prediction errors ($\text{UCE}=0.085 \pm 0.033$). This pipeline effectively segments complex histopathological structures and flags ambiguous regions for review, supporting downstream biomarker discovery and clinical interpretation.

Submission Number: 4