Improved U-Net based on contour prediction for efficient segmentation of rectal cancer.

BACKGROUND AND OBJECTIVE

Segmentation of rectal cancerous regions using 2D Magnetic Resonance Imaging (MRI) images is a critical step in radiation therapy. The shape of rectal cancer has significant variations and the shape of some surrounding organs is similar to that of rectal cancer; these conditions significantly affect the segmentation accuracy of rectal cancer and lead to incorrect segmentation. Therefore, automatic segmentation of rectal cancer is urgently needed, and it is a great challenge. For this task, the existing deep learning-based approaches have two shortcomings: 1) The U-Net network plays an important role in the field of medical segmentation. However, the designs of encoders and decoders in traditional U-Net networks are relatively simple and cannot extract good features, resulting in incorrect segmentation results. 2) Conventional neural networks extract high-level features that often do not include sufficient high-resolution contour information, resulting in ambiguity in contour segmentation. In this paper, we propose an improved U-Net network based on contour prediction, aiming at effective segmentation of rectal cancer.

METHODS

We designed a new U-Net network by improving the traditional U-Net network. We made four improvements: 1) We replaced the encoders with the SENet network. 2) A global pooling layer was added after the last encoder. 3) We added the Spatial and Channel Squeeze & Excitation (SCSE) attention mechanism module to each decoder. 4) We concatenated the output results of each decoder. In addition, the model implemented content segmentation and contour segmentation for rectal cancer in parallel, so that both the content and contour information was learned by the network to enhance the segmentation accuracy.

RESULTS

Our data were obtained from the Shanxi Provincial Cancer Hospital and included 3773 2D MRI rectal cancer images. The proposed method achieved an Mean Intersection over Union of 0.894 (MIoU) on the test set. Compared with state-of-the-art methods, our method had the best performance on the test set, and its MIoU metric was 0.123 higher than that of the second-best model. At the same time, the effectiveness of the improvements to our method was demonstrated through ablation experiments.

CONCLUSIONS

Our method can help radiologists to segment effectively, save their time and energy, and enable them to focus on cases that are not easily segmented because of the complex shape of rectal cancer.