Transformer-based weakly supervised intracerebral hemorrhage segmentation using image-level labels

Abstract

Objective

The segmentation of intracerebral hemorrhage (ICH) lesions in brain CT scans is of paramount importance for the diagnosis and treatment of stroke. Given the tremendous challenge of pixel-wise annotation in intracerebral hemorrhage, weakly supervised segmentation for ICH based on image-level labels has drawn great attention. Typical methods constructed based on convolutional neural networks often suffer from insufficient global perception, making it difficult to address ICH lesion diversity. Therefore, vision transformer, building pair-wise global dependency, becomes a popular alternative. Unfortunately, the data-hungry nature of vision transformer hinders its full exploitation given relatively limited medical imaging data, resulting in over-smoothing issue.

Methods

In this paper, based on the observation that most patches/tokens tend to build pair-wise dependency with intracerebral hemorrhage lesion, we propose weighted attention fusion (WAF) to fully utilize over-smoothing attention maps produced by ViT under conditions of limited training data. Compared to existing research, no additional parameters or computational complexity is introduced by WAF when incorporating target-relevant information. In addition, to recall low-confidence/-salient regions in segmentation, a patch-erasing re-activation mechanism is proposed by forcing the model to explore more class-specific regions.

Results

Experimental results on three datasets, i.e., INSTANCE2022, LocalBrainCT and BraTS2021 demonstrates the effectiveness of the proposed ICH weakly supervised segmentation framework. Compared to the previous works on the weakly supervised sementation, the proposed architecture obtains the state-of-the-art performance on intracerebral hemorrhage segmentation (Dice of 72.39).

Conclusion

This study focus on weakly supervised intracerebral hemorrhage segmentation, and propose a transformer-based framework with weighted attention fusion module and patch-erasing re-activation mechanism. It achives superior performance than previous methods under various settings.