T-stage diagnosis of lung cancer based on deep learning in CT images

Résumé fourni par la source

Objective: To explore the advantages of Swin-Transformer (SwinT) network in T-staging diagnosis of lung cancer computed tomography (CT) images by comparing it with Res-Net, Vgg-Net, and Mobile-Net networks. Methods: CT images of a total of 176 patients with lung cancer confirmed by pathological biopsy were collected from the First Affiliated Hospital of Army Medical University between 2021 and 2023. Based on the patients’ pathological T-staging reports, the patients’ images were classified into 4 categories: T1, T2, T3, and T4, and all the patients were simply randomly divided into a training set (n = 123), a validation set (n = 18), and a test set (n = 35) in the ratio of 7∶1∶2 for training the intelligent T-staging diagnostic model, which was used for training the intelligent T-staging diagnostic model by using the accuracy, precision, confusion matrix, recall rate, F1 score (F1- Score), receiver operating haracteristic (ROC) curve, and area under the ROC curve (AUC) parameters to assess the diagnostic efficacy of the network. Results: In the four classifications of T1, T2, T3, and T4, the accuracy in the training set of Res-Net, Vgg-Net, Mobile-Net, and SwinT network models were 0.5278, 0.6111, 0.6389, and 0.7222, as well as the AUC was 0.7275, 0.7850, 0.7275, and 0.8650.the SwinT network model had the best combined results. Conclusion: Compared with Res-Net, Vgg-Net and Mobile-Net networks, SwinT network achieves the optimal prediction performance in the lung cancer CT image classification task, and can be used for the smart T-staging of lung cancer to be able to diagnose, to improve the diagnosis and treatment efficiency, to shorten the diagnosis time, and to save medical resources.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.