Artificial intelligence and machine learning

Research Article

Integrating Multi-Scale Features and Attention Mechanisms for Colorectal Tumor Segmentation in CT Images

Yuqian Wang¹, Sergey Vladimirovich Aksenov²

1,2	School of Information Technologies and Robotics, Tomsk Polytechnic University, Tomsk, Russia
2	Tomsk State University of Control Systems and Radioelectronics, Tomsk, Russia
1	wangyuqian3333@gmail.com

Abstract. In recent years, deep learning technologies have been widely applied in medical image analysis, demonstrating outstanding performance particularly in image segmentation tasks.

To address the issue of semantic information loss during the feature extraction stage in U‑shaped networks, which limits the segmentation accuracy of colorectal tumors, this paper proposes a novel segmentation model based on the U‑Net architecture, named MGA‑UNet (Multi‑scale Ghost Attention U‑Net). The model integrates multi‑scale feature extraction, dual channel and spatial attention mechanisms, and attention gating in skip connections. The specific improvements are as follows:

First, an enhanced Ghost module (combined with RFB) is adopted in the encoding stage to achieve extraction and fusion of multi‑scale feature information.

Second, the CBAM (Convolutional Block Attention Module) is introduced into the encoding path to enhance the network's feature response to small‑scale targets.

Third, attention gate units are embedded in the skip connections to suppress irrelevant background regions and highlight tumor features.

Experimental results on a colorectal tumor CT dataset demonstrate the high effectiveness of the proposed model. Compared with the classic U‑Net, GhostNet, and the recent Mamba‑UNet and U‑SAM, the proposed model can delineate colorectal tumor regions more accurately and achieves superior segmentation performance. Furthermore, ablation studies and hyperparameter sensitivity analysis verify the effectiveness and stability of each proposed module. (Linked article texts in English and in Russian).

Keywords: U‑Net, attention mechanism, skip connection, image segmentation, MGA‑UNet

MSC-2020 68U10; 92C50

Acknowledgments: This work was supported by the China Scholarship Council (CSC) under grant No. 202008410491.

For citation: Yuqian Wang, Sergey V. Aksenov. Integrating Multi-Scale Features and Attention Mechanisms for Colorectal Tumor Segmentation in CT Images. Program Systems: Theory and Applications, 2026, 17:2, pp. 147–190. (in Engl. In Russ.). https://psta.psiras.ru/2026/2_147-190.

Full text of bilingual article (PDF): https://psta.psiras.ru/read/psta2026_2_147-190.pdf (Clicking on the flag in the header switches the page language).

The article was submitted 19.04.2026; approved after reviewing 28.04.2026; accepted for publication 29.05.2026; published online 20.06.2026.