Enhancing non-small cell lung cancer tumor segmentation with a novel two-step deep learning approach

Abstract

Precise recognition and delineation of tumors play a pivotal role in the radiotherapy of non-small cell lung cancer (NSCLC). The current manual delineation techniques used in clinical practice are both time-consuming and labor-intensive. While some studies have made progress in automating the segmentation of lung cancer targets, there is still room for further improvement in their effectiveness. The present study introduces a unique two-stage deep learning (DL) method. Computed tomography (CT) images of two hundred NSCLC cases are collected to train the network, forty cases' images are collected for validation and sixty cases’ images are collected for testing. A coarse network of segmentation is included in the first step, with a primary objective to detect the harsh region of the lesion. In this network, a competing manner between lesions and normal tissues is established based on the distribution of tissue structure, which will improve the ratio of the distinguished lesion overlapping the ground truth (GT) and decrease misidentification. The second step involves the utilization of two distinct segmentation networks for the two categories of images containing large-sized and small-sized tumors. These two networks are trained separately using the respective datasets, ultimately leading to the successful accomplishment of precise tumor delineation. Finally, the suggested approach produced a significantly higher dice similarity coefficient (DSC) of 0.80 ± 0.13 than other methods, achieving the highest TPR, the lowest FPR and HD95. In addition, our two-step method has greatly improved in both CT images with large and small GTVs. Among them, the performance is significantly advanced on CT images with small GTVs, which have poor performance in other methods. The study demonstrates that the suggested approach can precisely segment NSCLC tumors and have the potential to improve the radiotherapy efficiency.