Multi - scale context U-net for breast cancer postoperative radiotherapy in patients with brachial plexus

In this study, we propose a Multi-Scale Context U-net (MSC-U-net) network model for the precise and automated segmentation of the brachial plexus in patients who have undergone postoperative radiotherapy for breast cancer. A total of 389 patients who underwent postoperative radiotherapy for breast cancer were included in the training set, while 55 patients were included in the test set. The network model proposed in this study was optimized and trained to achieve the most accurate segmentation results. The performance of the model was evaluated using the Dice Similarity Coefficient (DSC) and the 95% Hausdorff distance (HD95). To further validate the effectiveness of the segmentation, comparison and ablation experiments were conducted. Subsequently, the clinical practicability was assessed within a clinical setting. MSC-U-net achieved precise segmentation results with DSC and HD95 values of 79.16 ± 0.05% and 6.95 ± 0.76 mm, respectively, in the test set. In comparison experiments with four other classical networks, the model showed significant statistical differences (p < 0.05) in performance. Ablation experiments further confirmed that the MSC-U-net network model achieved the best segmentation performance (p < 0.05). The radiation oncologists’ subjective evaluations also demonstrated the clinical applicability of the MSC-U-net network model. There was no statistically significant difference between manual segmentation and model segmentation in terms of segmentation accuracy and radiation dose (p > 0.05). The above results demonstrate the superior performance of the MSC-U-net network model in medical image segmentation, and also indicate its effectiveness in clinical applications. In the context of segmenting the brachial plexus in localization CT images of breast cancer patients after postoperative radiotherapy, MSC-U-net has demonstrated exceptional performance, significantly minimizing the manual segmentation accuracy issues caused by human factors. This network exhibits high efficiency in automatic segmentation and a high level of accuracy. Notably, the MSC-U-net network holds significant importance in the advancement of radiotherapy automation, and it also offers valuable insights for future research in the field of automatic organ segmentation. (1) The structure of Non-Local Block and SE Block is reconsidered, and Multi-Scale Context Block is proposed. The multi-scale self-attention mechanism is used to model the global context, and the MLP is used to extract important features to effectively improve the segmentation accuracy. (2) In this study, we employed the nearest neighbor interpolation upsampling layer to replace the traditional transposed convolution layer. This design effectively mitigates the common checkerboard artifacts issues during the upsampling process and enables the generation of smoother segmentation boundaries, thereby significantly improving segmentation accuracy. (3) The accuracy of automatic segmentation of the brachial plexus nerve in postoperative breast cancer radiotherapy patients was greatly improved by the proposed method.