Due to the increasing prevalence of lung cancer and the urgent need for artificial intelligence technologies to support medical fields, there is an immediate demand for highly accurate and dependable automated diagnostic models for the segmentation of chest CT scans. Accurate segmentation will reduce clinicians' manual labor and improve diagnostic results. This paper proposes a new segmentation model, which combines an improved U-Net architecture with a squeeze-and-excitation mechanism. The model can utilize the loss in a multilevel fashion, thanks to the auxiliary classification at the bottleneck layer. Furthermore, the squeeze-and-excitation block focuses on the key characteristics extracted from convolutional layers. Experiments on the LUNA16 dataset were carried out to train and evaluate the model. The proposed approach outperformed many other segmentation methods and was very effective in defining the shape of lung nodules more precisely. On the LUNA16 dataset, it achieved 98.42% DSC and 98.56% accuracy. It has been proven that the segmentation of lung nodules in chest CT scans using an auxiliary U-Net architecture together with a squeeze-and-excitation mechanism is highly efficient and cost-effective. High accuracy and Dice coefficient against the most advanced segmentation models demonstrate the possibility of clinical usage and provide accurate boundary detection, which is crucial for an accurate diagnosis and treatment. The proposed contraction and stimulation mechanism with an auxiliary U-Net architecture has achieved very effective yet inexpensive lung nodule segmentation on chest CT scans, compared to state-of-the-art segmentation models. Accurate identification of the boundaries is of paramount importance for diagnosis and treatment planning, and the high Dice coefficient and accuracy support its use in clinical settings.

LUNA16 Lung Nodules Segmentation Auxiliary Loss Squeeze and Excitation Block U-Net