Multi-classification-assisted diagnosis of multi-scale lung adenocarcinoma pathological maps based on Scalenet101

Abstract

Abstract Among all types of cancers, lung cancer is the deadliest cancer with a high mortality rate. Early diagnosis of lung cancer enables its timely and effective treatment, which can help reduce the risk of death. The aim of this research is to propose a deep learning approach for classifying and detecting lung adenocarcinoma tissues. The method was an improved Scalenet101 based on class activation mapping. (I) Scalenet101 classified lung adenocarcinoma tissue images into cancer, stromal and normal categories with the help of the Sigmoid. Then, the performance of Scalenet101 was validated using the pre-trained downstream techniques such as AlexNet, VGG16, VGG19, and ResNet50. (ii) Multi-scale fusion and weight fitting were performed to obtain multi-classification activation maps, which could improve the interpretability and accuracy in lung cancer evaluation. The performance of this model was tested using benchmark lung adenocarcinoma tissue images from WSSS4LUAD. The results showed that the model achieved an absolute classification accuracy of over 93.86% and an F1 score of over 95.44%.