ScribFormer: Transformer Makes CNN Work Better for Scribble-Based Medical Image Segmentation

Abstract

Most recent scribble-supervised segmentation methods commonly adopt a CNN framework with an encoder-decoder architecture. Despite its multiple benefits, this framework generally can only capture small-range feature dependency for the convolutional layer with the local receptive field, which makes it difficult to learn global shape information from the limited information provided by scribble annotations. To address this issue, this paper proposes a new CNN-Transformer hybrid solution for scribble-supervised medical image segmentation called ScribFormer. The proposed ScribFormer model has a triplebranch structure, i.e., the hybrid of a CNN branch, a Transformer branch, and an attention-guided class activation map (ACAM) branch. Specifically, the CNN branch collaborates with the Transformer branch to fuse the local features learned from CNN with the global representations obtained from Transformer, which can effectively overcome limitations of existing scribble-supervised segmentation methods. Furthermore, the ACAM branch assists in unifying the shallow convolution features and the deep convolution features to further improve model's performance. Extensive experiments on two public datasets and one private dataset show that our ScribFormer has superior performance over the state-of-the-art scribble-supervised segmentation methods, and achieves even better results than the fullysupervised segmentation methods. The code is released at https://github.com/HUANGLIZI/ScribFormer.

PIS-Net: Efficient Medical Image Segmentation Network with Multivariate Downsampling for Point-of-Care

“…The chosen methods are all considered classic and representative within the medical image segmentation field. They encompass convolutional baselines such as U-Net [14], U-Net++ [15], U-Net3+ [17] and ResU-Net [16]; recent transformer-based baselines such as TransU-Net [23] and MedT [24]; the CNN-Transformer hybrid model ScribFormer [29]; and the CNN-MLP hybrid model UNeXt [4].…”

“…LET-Net [27] combines a U-shaped CNN with a transformer effectively in a capsule embedding style to compensate for respective deficiencies; Yuan et al [28] proposed CTC-Net, which utilizes dual coding paths of CNNs and transformer encoders to produce complementary features. ScribFormer [29] improves model performance by utilizing a three-branch structure to unify shallow and deep features, which consists of a mixture of CNN branches, transformer branches, and attention-guided class activation map (ACAM) branches. CNN-MLP hybrid models' strength is that it can achieve better segmentation results while having a small number of parameters and fast reasoning.…”