Voxels Intersecting along Orthogonal Levels Attention U-Net for Intracerebral Haemorrhage Segmentation in Head CT

TL;DR

This paper introduces viola-Unet, a novel attention-based U-Net architecture that significantly improves intracranial hemorrhage segmentation in head CT scans, outperforming existing models in a competitive challenge.

Contribution

The paper presents viola-Unet, a new attention mechanism integrated into U-Net for better feature fusion, achieving state-of-the-art results in ICH segmentation.

Findings

Outperformed nnU-Net baseline in all metrics

Won the INSTANCE 2022 Data Challenge validation phase

Code and pretrained models are publicly available

Abstract

We propose a novel and flexible attention based U-Net architecture referred to as "Voxels-Intersecting Along Orthogonal Levels Attention U-Net" (viola-Unet), for intracranial hemorrhage (ICH) segmentation task in the INSTANCE 2022 Data Challenge on non-contrast computed tomography (CT). The performance of ICH segmentation was improved by efficiently incorporating fused spatially orthogonal and cross-channel features via our proposed Viola attention plugged into the U-Net decoding branches. The viola-Unet outperformed the strong baseline nnU-Net models during both 5-fold cross validation and online validation. Our solution was the winner of the challenge validation phase in terms of all four performance metrics (i.e., DSC, HD, NSD, and RVD). The code base, pretrained weights, and docker image of the viola-Unet AI tool are publicly available at \url{https://github.com/samleoqh/Viola-Unet}.