CLAHE-RATSNet: a deep learning framework for accurate lung nodule segmentation in computed tomography images

Abstract Lung nodule segmentation from computed tomography (CT) images is a crucial task for computer-aided diagnosis, treatment planning, and early diagnosis of lung cancer. Unfortunately, the segmentation of nodules is difficult in practice due to several factors, such as low contrast, small size, weak boundaries and the similarity of nodules to the vessels and the surrounding lung tissues. To solve these problems, this paper presents a click-guided lung nodule segmentation network that performs residual attention two-stage segmentation, called CLAHE-RATS-Net, which is based on the Contrast Limited Adaptive Histogram Equalization (CLAHE) method. The framework proposed here combines CLAHE with a Gaussian click map, which was generated from the centroid of the nodules, to get an enhanced CT contrast at the local level, and it also uses spatial guidance. A first-stage attention network is used for coarse localization, and a second-stage residual refinement module is used to correct the errors of boundaries with the improved ROI. Patient-wise data splitting was used to carry out the experiments on LIDC-IDRI CT images and the results were compared with baseline models. Under this oracle-guided test condition, the proposed model achieved a Dice score of 0.9238, an IoU of 0.8610, a precision of 0.9239, and a recall of 0.9238. Patient-level paired analyses showed statistically significant differences between CLAHE-RATSNet and the evaluated baseline models in Dice, IoU, precision, and recall after Holm correction for multiple comparisons. These findings characterize segmentation performance under the controlled oracle-guided evaluation protocol and should not be interpreted as end-to-end lesion-localization or fully automatic clinical performance.