EFCNet: Every Feature Counts for Small Medical Object Segmentation

TL;DR

A novel model named EFCNet is proposed for small object segmentation in medical images that addresses information loss during encoding and decoding procedures and significantly outperforms previous segmentation methods designed for both medical and normal images.

Abstract

This paper explores the segmentation of very small medical objects with significant clinical value. While Convolutional Neural Networks (CNNs), particularly UNet-like models, and recent Transformers have shown substantial progress in image segmentation, our empirical findings reveal their poor performance in segmenting the small medical objects and lesions concerned in this paper. This limitation may be attributed to information loss during their encoding and decoding process. In response to this challenge, we propose a novel model named EFCNet for small object segmentation in medical images. Our model incorporates two modules: the Cross-Stage Axial Attention Module (CSAA) and the Multi-Precision Supervision Module (MPS). These modules address information loss during encoding and decoding procedures, respectively. Specifically, CSAA integrates features from all stages of the encoder to adaptively learn suitable information needed in different decoding stages, thereby reducing information loss in the encoder. On the other hand, MPS introduces a novel multi-precision supervision mechanism to the decoder. This mechanism prioritizes attention to low-resolution features in the initial stages of the decoder, mitigating information loss caused by subsequent convolution and sampling processes and enhancing the model's global perception. We evaluate our model on two benchmark medical image datasets. The results demonstrate that EFCNet significantly outperforms previous segmentation methods designed for both medical and normal images.

Figures (4)

• Figure 1: Samples of small medical objects in two datasets (S-HRD and S-Polyp) in our work.
• Figure 2: Comparison of our method (d) against conventional encoder-decoder based method in previous works (a)-(c). (a)&(b) In traditional methods ronneberger2015u8589312, the single-stage features of the encoder and the corresponding single-stage features of the decoder are fused by concatenation or addition, with one segmentation head at the end of the decoder. (c) Some methods oktay2018attentionzhang2022isnet attempt to add attention mechanisms to the encoder, which however are limited to single-stage features. And only one segmentation head is adopted at the end of the decoder. (d) Our method aggregates all features in each stages of the encoder through CSAA to guide the decoding procedure. Multi-resolution features in each stages of the decoder are segmented with multi-precision by multiple segmentation heads through MPS.
• Figure 3: Structure of our proposed EFCNet. (a) Overview of our method, featuring an Encoder-Decoder architecture equipped with the Cross-Stage Axial Attention Module (CSAA) and Multi-Precision Supervision Module (MPS). (b) Details of our Cross-Stage Axial Attention Module (CSAA). The CSAA combines features from each stage of the encoder, dynamically extracts information about small medical objects, and directs the decoding process of each stage in the decoder.
• Figure 4: Visualization of EFCNet (ours) and previous SOTA methods on S-HRD and S-Polyp. The previous SOTA method on S-HRD is Attn-UNet oktay2018attention, and the previous SOTA method on S-Polyp is SSFormer wang2022stepwise. The green circle areas show extremely small medical objects captured by our method that are not captured by previous SOTA methods. The yellow circle areas show that the segmentation of the boundaries of small medical objects in our method is significantly better than the previous SOTA method. The red circle areas show the wrong segmentation of small medical objects in the previous SOTA method while our method is correct.