Two heads are enough: DualU-Net, a fast and efficient architecture for cell classification and segmentation

Abstract

Accurate detection and classification of cell nuclei in histopathological images are critical for both clinical diagnostics and large-scale digital pathology workflows. In this work, we introduce DualU-Net, a fully convolutional, multi-task architecture designed to streamline cell nuclei classification and segmentation. Unlike the widely adopted three-decoder paradigm of HoVer-Net, DualU-Net employs only two output heads: a segmentation decoder that predicts pixel-wise classification maps and a detection decoder that estimates Gaussian-based centroid density maps. By leveraging these two outputs, our model effectively reconstructs instance-level segmentations. The proposed architecture results in significantly faster inference, reducing processing time by up to ×5 compared to HoVerNet, while achieving classification and detection performance comparable to state-of-theart models. Additionally, our approach demonstrates greater computational efficiency than CellViT and NuLite. We further show that DualU-Net is more robust to staining variations, a common challenge in digital pathology workflows. The model has been successfully deployed in clinical settings as part of the DigiPatICS initiative, operating across eight hospitals within the Institut Català de la Salut (ICS) network, highlighting the practical viability of DualU-Net as an efficient and scalable solution for nuclei segmentation and classification in real-world pathology applications. The code and pretrained model weights are publicly available on https://github.com/davidanglada/DualU-Net.

This publication is part of the R&D&I project PID2023-148614OB-I00, funded by MICIU/AEI/10.13039/501100011033/ and by FEDER, EU. This research has also been funded by European Development Funds Regional, Programa operatiu FEDER de Catalunya 20142020 through the project DigiPatICS.

Peer Reviewed

Postprint (published version)