Abstract:Untreated colorectal polyps can develop into colorectal cancer, which is a leading cause of cancer-related deaths. Colonoscopy is a commonly-used method for colorectal polyp scanning, but limited to the experience and subjectivity of clinicians, one out of four polyps cannot be correctly recognized. In this article, we propose an automatic colorectal polyp segmentation system based on the deep convolutional neural network, aiming to improve the accuracy of colorectal polyp scanning. The proposed ABC-Net is comprised of a shared encoder and two novel mutually-constrained decoders for simultaneous polyp area and boundary segmentation. To sufficiently exploit multi-scale image information, the selective feature modules are embedded into the network and used for dynamically learning and fusing multi-scale feature representations. Furthermore, a new boundary-sensitive loss is proposed to model the interdependencies between the area and boundary branches, the information of the two branches are reciprocally propagated and constrained, yielding a significant improvement in segmentation accuracy. Extensive experiments are conducted on three public colorectal polyp datasets, and the results, e.g., F1 scores are 0.866, 0.915, 0.874 in EndoScene, Kvasir-SEG, and ETIS-Larib datasets, demonstrate the advantages of the proposed method.

What problem does this paper attempt to address?

The problem that this paper attempts to solve is to improve the accuracy in colorectal polyp screening. Specifically, the paper proposes an automatic colorectal polyp segmentation system based on deep convolutional neural networks, aiming to improve the accuracy of colorectal polyp scanning. The paper mentions that due to the limitations of doctors' experience and subjective judgment, approximately 1/4 of polyps cannot be correctly identified during routine colonoscopy. Therefore, the goal of this study is to develop a system that can automatically and accurately segment colorectal polyps, so as to reduce medical errors and relieve the workload of doctors. ### Main contributions of the paper 1. **Proposing a new deep - learning network**: This network consists of a shared encoder and two mutually - associated decoders, which are respectively used for simultaneously segmenting the polyp area and the boundary. Selective Kernel Modules (SKMs) are embedded in the network to dynamically learn multi - scale feature representations. 2. **Designing a boundary - sensitive loss function**: This loss function utilizes region - boundary constraints, and the experimental results prove that the two branches can improve each other, thus producing more accurate polyp predictions. 3. **Verifying the effectiveness of the method on three public datasets**: Extensive experiments show that this method has achieved satisfactory performance in the colorectal polyp segmentation task. For example, the F1 - scores on the EndoScene, Kvasir - SEG and ETIS - Larib datasets are 0.866, 0.915 and 0.874 respectively. ### Method overview - **Network architecture**: The network consists of a shared encoder (SE - ResNeXt - 50) and two mutually - associated decoders (region branch and boundary branch). The encoder is responsible for extracting low - level features (such as texture, edge) and high - level features (such as shape, object), while the decoder reversely maps these features to match the size of the input image and predicts the polyp area and boundary respectively. - **Selective Feature Modules (SKMs)**: Embedded between adjacent decoding blocks, they are used to dynamically learn multi - scale feature representations, thereby enhancing the feature - expressing ability of the network. - **Boundary - sensitive loss function**: It consists of three parts: - **Region branch loss \(L_a\)**: It combines Dice loss and binary cross - entropy loss. - **Boundary branch loss \(L_b\)**: It minimizes the difference between the predicted boundary and the real boundary. - **Region - boundary constraint losses \(L_{C1}\) and \(L_{C2}\)**: The interdependence between the region and the boundary is modeled through two parts. Among them, \(L_{C1}\) measures the difference between the output of the contour detector and the real boundary, and \(L_{C2}\) makes the prediction results of the region and boundary branches consistent through the Kullback - Leibler divergence. ### Experimental results - **Datasets**: The experiments were carried out on three public datasets, namely EndoScene, Kvasir - SEG and ETIS - Larib. - **Performance indicators**: On these datasets, the F1 - scores of this method are 0.866, 0.915 and 0.874 respectively, showing its superior performance in the colorectal polyp segmentation task. In conclusion, by proposing a novel network architecture and loss function, this paper significantly improves the accuracy of colorectal polyp segmentation and provides strong support for clinical applications.

ABC-Net: Area-Boundary Constraint Network With Dynamical Feature Selection for Colorectal Polyp Segmentation

Boundary Constraint Network With Cross Layer Feature Integration for Polyp Segmentation

Boundary Refinement Network for Colorectal Polyp Segmentation in Colonoscopy Images

An Edge-Enhanced Network for Polyp Segmentation

Colorectal Polyp Segmentation Using a Fully Convolutional Neural Network

BLE-Net: Boundary Learning and Enhancement Network for Polyp Segmentation

ECTransNet: An Automatic Polyp Segmentation Network Based on Multi-scale Edge Complementary

Colorectal Polyp Segmentation by U-Net with Dilation Convolution

Automatic Polyp Segmentation in Colonoscopy Images Using a Modified Deep Convolutional Encoder-Decoder Architecture

Progressive Feature Enhancement Network for Automated Colorectal Polyp Segmentation

Accurate Polyp Segmentation Through Enhancing Feature Fusion and Boosting Boundary Performance.

DBE-Net: Dual Boundary-Guided Attention Exploration Network for Polyp Segmentation

Automatic polyp detection and segmentation using shuffle efficient channel attention network

Automatic Polyp Segmentation Via Multi-scale Subtraction Network

Dual-branch multi-information aggregation network with transformer and convolution for polyp segmentation

MCSF-Net: a multi-scale channel spatial fusion network for real-time polyp segmentation

Dual encoder–decoder-based deep polyp segmentation network for colonoscopy images

ICBNet: Iterative Context-Boundary Feedback Network for Polyp Segmentation

C 2 E-Net: Cascade attention and context-aware cross-level fusion network via edge learning guidance for polyp segmentation

IDDF2018-ABS-0260 Deep Learning for Polyp Segmentation

APRNet: Alternative Prediction Refinement Network for Polyp Segmentation.