[Paper Review] Adaptive Context Selection for Polyp Segmentation

Ruifei Zhang, Guanbin Li+4 more|arXiv (Cornell University)|Jan 12, 2023

Radiomics and Machine Learning in Medical Imaging12 citations

TL;DR

The paper proposes ACSNet, an encoder–decoder framework with Local Context Attention (LCA), Global Context Module (GCM), and Adaptive Selection Module (ASM) to adaptively fuse local and global context for polyp segmentation, achieving state-of-the-art results on EndoScene and Kvasir-SEG datasets.

ABSTRACT

Accurate polyp segmentation is of great significance for the diagnosis and treatment of colorectal cancer. However, it has always been very challenging due to the diverse shape and size of polyp. In recent years, state-of-the-art methods have achieved significant breakthroughs in this task with the help of deep convolutional neural networks. However, few algorithms explicitly consider the impact of the size and shape of the polyp and the complex spatial context on the segmentation performance, which results in the algorithms still being powerless for complex samples. In fact, segmentation of polyps of different sizes relies on different local and global contextual information for regional contrast reasoning. To tackle these issues, we propose an adaptive context selection based encoder-decoder framework which is composed of Local Context Attention (LCA) module, Global Context Module (GCM) and Adaptive Selection Module (ASM). Specifically, LCA modules deliver local context features from encoder layers to decoder layers, enhancing the attention to the hard region which is determined by the prediction map of previous layer. GCM aims to further explore the global context features and send to the decoder layers. ASM is used for adaptive selection and aggregation of context features through channel-wise attention. Our proposed approach is evaluated on the EndoScene and Kvasir-SEG Datasets, and shows outstanding performance compared with other state-of-the-art methods. The code is available at https://github.com/ReaFly/ACSNet.

Motivation & Objective

• Motivate accurate polyp segmentation by addressing polyp size and shape diversity.
• Develop an encoder–decoder framework that selectively leverages local and global context.
• Introduce modules (LCA, GCM, ASM) to adaptively fuse contextual features based on polyp size.
• Evaluate on public datasets (EndoScene, Kvasir-SEG) and show improved performance over state-of-the-art.

Proposed method

• Enhanced UNet-like encoder–decoder with five-block ResNet34 encoder/decoder.
• Local Context Attention (LCA) modules replace skip connections to enrich decoder features with hard-region local context guided by the previous prediction map.
• Global Context Module (GCM) captures multi-scale global context via pooling branches and non-local operation, feeding context to all ASM blocks.
• Adaptive Selection Module (ASM) uses squeeze-and-excitation style channel-wise attention to fuse local and global context with previous decoder features.
• Prediction maps at multiple decoder stages supervise learning; loss combines binary cross-entropy and dice loss.

Experimental results

Research questions

• RQ1 Can adaptive selection of local and global context improve polyp segmentation across variable polyp sizes and shapes?
• RQ2 How do LCA, GCM, and ASM contribute individually and collectively to segmentation accuracy?
• RQ3 Does ACSNet achieve consistent improvements over existing architectures on standard benchmarks?

Key findings

• ACSNet achieves Dice score of 86.59% on EndoScene, the best among listed methods, with a 3.66% improvement over the second-best.
• On EndoScene, ACSNet also achieves Rec 87.96, Spec 99.16, Prec 90.99, IoUp 79.73, IoUb 96.86, mIoU 88.29, Acc 97.11.
• On Kvasir-SEG, ACSNet achieves Dice 91.30 and mIoU 90.28, outperforming UNet, UNet++, SFANet, and SegNet variants; Baseline+LCAs+GCM and other ablations also show progressive gains.
• Ablation studies indicate adding LCAs, GCM, and ASM incrementally improves Dice by 0.79%, 1.28%, and 1.02% respectively over baseline.