Transformer-based Lung Cancer Segmentation Using Dual Window Images in Chest CT Images

Lung cancer varies in size, shape, and location. Furthermore, isolated tumors located in the lung parenchyma have clear boundaries, whereas tumors attached to surrounding structures with similar intensity values have indistinct boundaries, making tumor segmentation challenging. To address these issues, we propose a transformer-based lung cancer segmentation method using dual-window images to resolve the boundary ambiguity of lung cancers attached to structures such as the mediastinum and chest wall in chest CT images. To effectively represent the characteristics of lung cancer, we generate lung window and mediastinal window images using two window settings. The proposed method utilizes SegFormer as the backbone and employs a dual-encoder architecture, where each image is input into a different encoder to generate multi-scale feature maps. In the decoder, feature map concatenation and channel attention blocks are applied to assign weights to the key features. Through experiments, the proposed method demonstrated the best performance in overall lung cancer segmentation, achieving a DSC of 75.74% and a recall of 80.44%. Additionally, in the segmentation results by the location of the lung cancer, the proposed method achieves the highest recall.