Fast and accurate lung cancer subtype classication and localization based on Intraoperative frozen sections of lung adenocarcinoma.

Abstract

Objective.Current lung cancer diagnostic techniques primarily focus on tissue subtype classification, yet remain inadequate in distinguishing pathological progression subtypes (particularly between adenocarcinomain situand invasive adenocarcinoma) on frozen sections. This study develops a deep neural network-based auxiliary diagnostic system specifically for surgical frozen sections, aiming to reduce pathologists' diagnostic workload while improving differentiation accuracy.Approach.We developed an innovative deep learning system (FSG-TL Model) for lung adenocarcinoma frozen section analysis, combining multi-instance learning with EMA/SimAM/SE attention-enhanced ResSimAM_Hybrid model for classification. Create carefully annotated frozen section datasets. FSG-TL Model integrates down sampling, tissue localization and classification to achieve automatic cancer detection, and improves classification performance through image enhancement and classification model optimization.MainResults.The method developed in this study exhibited significant accuracy in identifying cancerous regions in frozen sections while successfully distinguishing between various cancer subtypes. A comprehensive automated localization system for lung adenocarcinoma full-scan sections was adeptly constructed, enabling swift localization of a 40,000×60,000 pixel full slide image in around 3 minutes. Notably, in the subtype instance classification of tumor region localization, ResSimAM_Hybrid achieved a classification accuracy (ACC) of 90.72%, outperforming the computational-pathology foundation model UNI. For the tumor localization task, the FSG-TL Model attained a tumor localization Dice score of 0.82. The localization Dice score for AIS and IAC reached 0.77 and 0.69, respectively.Significance.This study provides a fast and accurate method for localizing cancer and lung adenocarcinoma subtypes in frozen sections. It provides important support for future research on AI-assisted clinical diagnosis of lung adenocarcinoma in frozen sections and reveals the research potential of AI-assisted diagnosis of subtypes of lung adenocarcinoma in the stage of pathological progression.