基于深度学习的肺腺癌气腔播散病灶识别和半定量评估模型的构建（英文）

【作者】 丁瀚林； 冯一鹏； 黄兴； 夏文杰； 许林； 董高超； 蒋峰；

【机构】 江苏省肿瘤医院/江苏省肿瘤防治研究所/南京医科大学附属肿瘤医院；

【摘要】 Purpose: Since 2015, the World Health Organization classified Spread Through Air Space(STAS)as a new invasion mode of lung cancer. Medical researchers gradually realized that STAS has its ow potential value in clinical treatment decision making. The reported incidence of STAS in NSCIC varies from 16% to 60% in multiple studies conducted worldwide. The detection rate of STAS exhibits significant differences among doctors, primarily influenced by their experience.Methods: 506 digital WSIs of 285 LUAD patients were collected. A STAS detection model,named STASNet, was constructed based on the MobileNetV3 model, calculating semi-quantitative parameters associated with the density and distance of STAS to predict patient recurrence. The artificial intelligence(AI)-assistance workflow was established to assist the STAS detection and assessment of recurrence risk.Results: The STASNet exhibited an accurate rate of 0.93 for STAS detection on the tiles level and had an AUC ranging from 0.72-0.78 for determining the STAS status across three datasets on the whole slide images(WSI) level. Among the semi-quantitative parameters, T10S, combined with spatial location information, which significantly stratified stage I LUAD patients on the disease-free survival(DFS). As for the lightweight architecture of MobileNetV3, we also deployed the STASNet into a real-time pathological diagnostic environment, which boosted the detection rate of STAS and identified three easily misidentified types of occult STAS.Conclusion: he deep learning mode established in this study showed the ability of predicting STAS lesions on tumor boundary image and the WSI level image of lung cancer. The model can also help pathologists to focus on the high incidence area of STAS in WSls of lung cancer, which can improve reading efficiency. At the same time, this model can reduce cross observer bias among clinical physicians and improve diagnostic consistency for STAS.更多还原

【Abstract】 Purpose: Since 2015, the World Health Organization classified Spread Through Air Space(STAS)as a new invasion mode of lung cancer. Medical researchers gradually realized that STAS has its ow potential value in clinical treatment decision making. The reported incidence of STAS in NSCIC varies from 16% to 60% in multiple studies conducted worldwide. The detection rate of STAS exhibits significant differences among doctors, primarily influenced by their experience.Methods: 506 digital WSIs of 285 LUAD patients were collected. A STAS detection model,named STASNet, was constructed based on the MobileNetV3 model, calculating semi-quantitative parameters associated with the density and distance of STAS to predict patient recurrence. The artificial intelligence(AI)-assistance workflow was established to assist the STAS detection and assessment of recurrence risk.Results: The STASNet exhibited an accurate rate of 0.93 for STAS detection on the tiles level and had an AUC ranging from 0.72-0.78 for determining the STAS status across three datasets on the whole slide images(WSI) level. Among the semi-quantitative parameters, T10S, combined with spatial location information, which significantly stratified stage I LUAD patients on the disease-free survival(DFS). As for the lightweight architecture of MobileNetV3, we also deployed the STASNet into a real-time pathological diagnostic environment, which boosted the detection rate of STAS and identified three easily misidentified types of occult STAS.Conclusion: he deep learning mode established in this study showed the ability of predicting STAS lesions on tumor boundary image and the WSI level image of lung cancer. The model can also help pathologists to focus on the high incidence area of STAS in WSls of lung cancer, which can improve reading efficiency. At the same time, this model can reduce cross observer bias among clinical physicians and improve diagnostic consistency for STAS.更多还原