Study On Deep Learning Method For Subtype Classification Of Non-small Cell Lung Cancer Based On CT

Non-small cell lung cancer is a serious threat to human life and health.There are significant differences between the two most common subtypes of adenocarcinoma and squamous cell carcinoma in terms of treatment options and survival time.Using CT images to classify lung cancer subtypes can help doctors formulate targeted treatment plans,thereby extending the survival time of patients,and has great research significance and application value.In recent years,deep learning technology has been effectively applied to lung cancer subtype classification tasks,but how to fully and efficiently extract tumor features from CT images and reduce background noise interference remains to be solved.Therefore,this article proposes a variety of novel deep learning methods to deeply mine the rich tumor morphological and structural information contained in multiple views and slices of CT images as well as the spatial correlation information contained in continuous slices,while effectively separating tumor information and background noise,thereby achieving a more accurate classification of non-small cell lung cancer subtypes.The specific research content is as follows:(1)Aiming at the problems of insufficient utilization of tumor information and interference from background information,we propose a novel method for subtype classification of non-small cell cancer based on multi-view learning and feature decomposition,MVFD-Net.This method uses tumor images from axial,sagittal,and coronal views of CT images as multi-view inputs to comprehensively consider the tumor information contained in different views.In order to separate background information from features of multiple views,a feature decomposition module is designed to decompose common features representing tumor information and specific features representing background environment information through attention mechanisms.A multi-view cross reconstruction loss function is proposed to ensure the effectiveness of feature decomposition,and a multi-view feature similarity loss function is proposed to further enhance the similarity of common features in multiple views.This article compares the classification performance of MVFD-Net and existing methods,and the results show that this method has excellent classification ability for non-small cell lung cancer subtypes,and its classification performance is superior to other methods.(2)In order to fully mine the tumor morphological structure information contained in different views and slices of CT images,based on the above research,we further propose MVMS-Net,a non-small cell lung cancer subtype classification method using multi-view and multi-slice of CT images.This method first extracts features from tumor center slices and multiple adjacent slices from multiple views of CT images,and separates background information from different slices through feature decomposition and multiple loss functions.Later,in order to fully capture the spatial correlation information contained in consecutive slices,this paper also designed a weighted multislice fusion module based on recurrent neural networks to facilitate network learning of more discriminative feature representations.The experimental results show that MVMS-Net can accurately distinguish between adenocarcinoma and squamous cell carcinoma,further improving the classification performance of non-small cell lung cancer subtypes.