ECHO: A unified framework for efficient cell-graph construction for breast histology classification

Background: The growing complexity of oncological diagnostics demands automated techniques to support pathologists in precise cancer subtype identification. There is a critical need for analytical tools that streamline analysis and compress large datasets into compact representations, optimizing storage resources. Among various methodologies, graph-based representations have demonstrated efficacy in classifying pathological images. Method: We propose a novel approach that streamlines cell-graph construction through a unified framework. Unlike conventional methods relying on two separate networks for nuclei detection and feature extraction, our approach leverages a single network for both tasks. Specifically, we employ HoVer-UNet, a neural network specialized in nuclei instance segmentation and classification, optimized as a distilled version of HoVer-Net. Results: Our approach reduces cell-feature dimensionality and graph storage footprint. Empirical evaluations on the BRACS benchmark yield a weighted F1 of 59.6% in the 7-class breast cancer subtype classification setting, surpassing the CG-GNN baseline by 3.7 percentage points. The proposed representation reduces per-node feature dimensionality by approximately 28 × (1 × 18 vs. 1 × 512) and overall graph storage footprint by up to 30 × . HoVer-UNet was previously reported to achieve approximately 3 × faster inference than HoVer-Net. Conclusions: The proposed method offers a streamlined, storage-efficient, and effective framework for cell-graph construction, with the potential to improve diagnostic workflows and support pathologists in precise cancer subtype identification.