LoGo Transformer: Hierarchy Lightweight Full Self-Attention Network for Corneal Endothelial Cell Segmentation

Engineering notes

Recently, Transformer outperforms convolution in modeling long-range dependencies but lacks local inductive bias so the pure transformer network is difficult to train on small medical image datasets. Compared with the convolution neural networks (CNNs) and the hybrid CNN-Transformer state-of-the-art (SOTA) methods, the LoGo transformer obtains the best result.

Chinese explanation / 中文解读

中文解读待补充：本站会优先为端到端自动驾驶、BEV感知、3D目标检测、轨迹预测、路径规划、LiDAR感知等高价值论文补充中文说明。

Original abstract

Corneal endothelial cell segmentation plays an important role in quantifying clinical indicators for the cornea health state evaluation. Although Convolution Neural Networks (CNNs) are widely used for medical image segmentation, their receptive fields are limited. Recently, Transformer outperforms convolution in modeling long-range dependencies but lacks local inductive bias so the pure transformer network is difficult to train on small medical image datasets. Moreover, Transformer networks cannot be effectively adopted for secular microscopes as they are parameter-heavy and computationally complex. To this end, we find that appropriately limiting attention spans and modeling information at different granularity can introduce local constraints and enhance attention representations. This paper explores a hierarchy full self-attention lightweight network for medical image segmentation, using Local and Global (LoGo) transformers to separately model attention representation at low-level and high-level layers. Specifically, the local efficient transformer (LoTr) layer is employed to decompose features into finer-grained elements to model local attention representation, while the global axial transformer (GoTr) is utilized to build long-range dependencies across the entire feature space. With this hierarchy structure, we gradually aggregate the semantic features from different levels efficiently. Experiment results on segmentation tasks of the corneal endothelial cell, the ciliary body, and the liver prove the accuracy, effectiveness, and robustness of our method. Compared with the convolution neural networks (CNNs) and the hybrid CNN-Transformer state-of-the-art (SOTA) methods, the LoGo transformer obtains the best result.

Links and sources

• Official / arXiv page

Comments