An Improved TransUNet Network for Lung Nodule Segmentation Mask Annotation

Lung cancer, one of the foremost causes of cancerrelated deaths, often presents without early symptoms, resulting in late-stage diagnoses. Radiotherapy’s efficacy depends on precise lung nodule segmentation to accurately locate tumor targets and avoid damaging organs at risk (OAR). Manual segmentation is time-consuming and varies by physician experience. While automatic segmentation offers efficiency and accuracy, it is sensitive to image quality and requires substantial labeled data for training. This paper introduces an interactive pulmonary nodule segmentation tool based on an improved TransUNet model. Key improvements include converting CNN layers to dilated convolutions to expand the receptive field and adding CBAM modules to enhance feature representation through channel and spatial attention mechanisms. The tool combines deep learning with manual intervention, allowing doctors to automatically segment lung nodules and manually refine results via an interactive interface, supporting polygon labeling, mask drawing, and user-defined editing.Results from experiments on the LIDC-IDRI dataset indicate that the improved model outperforms existing approaches, achieving Dice coefficient, IoU, and recall improvements of $\mathbf{0. 4 \%, 0. 6 \%}$, and $0.1 \%$ over the original TransUNet. Ablation studies confirm the effectiveness of dilated convolutions and CBAM modules in enhancing segmentation performance. This tool significantly improves segmentation accuracy and efficiency, providing better support for clinical diagnosis.