Domain and Content Adaptive Convolutions for Cross-Domain Adenocarcinoma Segmentation

TL;DR

The paper tackles cross-domain adenocarcinoma segmentation in histology under domain shifts caused by organ morphology and scanner differences. It builds on nnU-Net and integrates the dcac (domain- and content-adaptive convolution) module to enhance cross-domain generalization. Evaluations on the COSAS datasets show strong performance, with the final challenge rankings placing the nnU-Net baseline first in both tracks (Dice/Jaccard–based scores of 0.8020 for cross-organ and 0.8527 for cross-scanner), while the dcac variant demonstrates potential advantages in cross-domain transfer. The work advances robust, domain-aware segmentation in digital pathology and provides open-source tooling for reproducibility and further research.

Abstract

Recent advances in computer-aided diagnosis for histopathology have been largely driven by the use of deep learning models for automated image analysis. While these networks can perform on par with medical experts, their performance can be impeded by out-of-distribution data. The Cross-Organ and Cross-Scanner Adenocarcinoma Segmentation (COSAS) challenge aimed to address the task of cross-domain adenocarcinoma segmentation in the presence of morphological and scanner-induced domain shifts. In this paper, we present a U-Net-based segmentation framework designed to tackle this challenge. Our approach achieved segmentation scores of 0.8020 for the cross-organ track and 0.8527 for the cross-scanner track on the final challenge test sets, ranking it the best-performing submission.

Figures (1)

• Figure 1: Overview of the dcac module. Figure adapted from Hu et al.hu2022domain.