Abstract: Due to the irregularity of colorectal tumor contours, it is a challenging task for clinicians to segment colorectal tumors manually in CT scans. To solve this problem, a novel algorithm based on a convolutional neural network with an attention mechanism is proposed to automatically achieve tumor segmentation. The proposed network consists of three major modules: an encoder module, which is fed CT scans to attain the feature map; a dual attention module, which includes a channel attention module and a position attention module to obtain more contextual information in the deep layer of the network; and a decoder module, which restores the feature map to the original size of the input images. We used 1131 CT slices of colorectal tumors to train and test the proposed network. Compared with U-Net and CE-Net, the Dice coefficient increased by 1.46% and 0.66% respectively, for our model. The comprehensive results show that the proposed network performs more effectively in colorectal tumor segmentation than the other methods.
Introduction
Colorectal cancer (CRC) remains a leading cause of cancer-related mortality worldwide. Computed Tomography (CT) imaging serves as the primary tool for staging and surgical planning. However, manual segmentation of colorectal tumors is a labor-intensive process prone to high inter-observer variability. The primary challenges stem from the highly irregular shapes of the tumors and their density, which often blends seamlessly into adjacent healthy tissues.
While deep learning models have shown promise in medical imaging, standard architectures often struggle with boundary precision. Specifically, the frequent use of max-pooling layers, while beneficial for learning abstract features, tends to discard fine-grained spatial information necessary for defining precise tumor edges. This paper proposes a modified U-Net architecture incorporating a Dual Attention Module (DAM) to address these limitations.
Proposed Methodology
The proposed architecture refines the traditional U-Net structure through three strategic modifications designed to preserve spatial resolution and enhance feature discrimination.
1. Modified Encoder
The encoder utilizes four convolutional blocks. Critically, the authors removed the fourth max-pooling layer. By maintaining a higher resolution in the deeper layers of the network, the model preserves the structural "profile" of the lesion that is typically lost during aggressive downsampling.
2. Dual Attention Module (DAM)
Integrated between the encoder and decoder, the DAM acts as the network's cognitive filter. It consists of two parallel components:
- Position Attention Module (PAM): This module learns spatial dependencies by calculating the relationship between any two pixels, regardless of their distance. It effectively captures long-range contextual information to determine which spatial regions are most likely to belong to the tumor.
- Channel Attention Module (CAM): This module focuses on the "what" by assigning weights to different feature channels. It emphasizes the most discriminative feature maps while suppressing noise from irrelevant tissues.
3. Decoder and Feature Fusion
The decoder restores the feature map to the original input dimensions. Shortcut connections are employed to fuse low-level edge details from the encoder with the high-level semantic features refined by the DAM, ensuring that both global context and local boundaries are represented.
Experimental Setup
The study utilized a private dataset from the First Hospital of Jilin University, comprising 1,131 CT slices from 158 patients. Images were reconstructed into a 512 × 512 matrix. Ground truth masks were established through a rigorous process involving two experienced radiologists and verification by a chief physician with 25 years of expertise. The network was optimized using a Dice coefficient loss function to mitigate the class imbalance between the small tumor regions and the vast background pixels.
Results and Discussion
The proposed model demonstrated superior performance when compared to the classic U-Net and the more modern CE-Net. Quantitative improvements were observed across all primary metrics:
| Metric | Proposed Model | vs. U-Net (Gain) | vs. CE-Net (Gain) |
|---|---|---|---|
| Dice Coefficient | 0.7089 | +1.46% | +0.66% |
| Jaccard Index | 0.5782 | +1.60% | +0.92% |
| Recall | 0.7885 | +2.38% | +3.74% |
| F1-Score | 0.7089 | +1.15% | - |
Qualitative analysis revealed that the attention mechanism effectively acted as a noise filter. While U-Net frequently over-predicted tumor boundaries (flagging healthy tissue) and CE-Net occasionally under-predicted irregular edges, the proposed model produced contours that closely aligned with expert labels. The integration of PAM and CAM allowed the network to ignore tissues with similar densities to the tumor, focusing instead on the core lesion.
Conclusion
This research presents an effective automated solution for colorectal tumor segmentation. By combining a resolution-preserving encoder with a Dual Attention Module, the model achieves higher accuracy and smoother contours than standard benchmarks. Future work will focus on expanding the dataset through multi-center studies and exploring 3D convolutional neural networks to capture the volumetric context between adjacent CT slices.
Summary updated