A Symmetric Encoder-Decoder Network with Enhanced Group-Shuffle Modules for Robust Lung Nodule Detection in CT Scans.

Lung cancer is considered to be a significant cause of death in the world, and the timely identification of nodules in the lungs in CT scans is very important to enhance the prognosis of patients. Although the state of the art of nodule delineation using deep learning-based segmentation models was achieved, major problems, including high feature diversity, low spatial discrimination, and overfitting of the models, require stronger feature-processing approaches. This research explores an enhanced symmetric encoder-decoder segmentation network known as the Improved Group-Shuffle Module (IGSM) to overcome these shortcomings. The most important feature of the proposed method is the IGSM, which hierarchically divides feature maps into a few groups, then transforms them independently, and then randomly switches channels between groups to increase inter-group interaction of features and diversity. This IGSM method is inspired by human brain functions, which are processed in specialized cortex areas, which are mimicked in this work through small-group feature processing. Channel shuffling is designed based on inter-modular communication in the human brain through coherent information sharing among the small groups of cortices. Through this mechanism, the model is much better at capturing discriminative spatial and contextual patterns, especially on complex and subtle nodule structures. The IGSM configurations have been optimized, specifically, the placement of the modules, grouping size, and shuffle permutation strategies. The proposed model's performance is then compared with the benchmarked models, like U-Net and DeepLab, with various performance indicators such as mean Intersection over Union (mIoU), Dice Score, Accuracy, Sensitivity, and Specificity. The simulation results proved the superiority of the IGSM-enhanced model with the mIoU of 0.7735, the Dice Score of 0.9665, and the Accuracy of 0.9873. The addition of the group and shuffle module not only enhances the discrimination between the nodules and their background, but it also improves the ability to generalize over a variety of nodules' morphology, thus producing a reliable tool for automated detection of lung cancer.