Deep neural networks to predict pathological invasiveness through surgical resection images: A computer vision based prospective clinical trial in pulmonary nodules.

• DOI: 10.1200/jco.2023.41.16_suppl.3068
• Publication Date: 2023-06-04T15:23:35Z
• Authors: Henian Yao; Lintong Yao; HaiYu Zhou; Weijie Zhan; Xie Daipeng; Zekai Huang; Hongwei Lin
• Keywords: Nodule (geology)

3068 Background: Pathological subtypes associate with surgical strategies of pulmonary nodule, during which thoracic surgeons make decision through intraoperative frozen resection diagnosis. However, section is time-consuming, vague and maybe misleading due to the limited sample subjective judgements from pathologists. This study developed a predictive deep neural network for pathological invasiveness nodules based on images gross specimens. Methods: We prospectively collected specimen image under standardized lighting conditions in operating theaters patients treated by surgery June 2020 September 2021 Guangdong Provincial People’s Hospital. Images were assigned into training cohort, validating cohort test as 8:1:1. With data augmentation, DenseNet was applied classify high-risk low-risk group. High-risk group defined grade 2 3 invasive adenocarcinoma according IASLC, while situ (AIS), minimally (MIA) 1 adenocarcinoma. Predictive efficiency evaluated area curve (AUC) receiver characteristic (ROC) curve, sensitivity, specificity, accuracy. Model performance would be compared Results: Among 1022 enrolled, we have acquired 1080 resected Hospital, mean age 56 years 39.1% male patients. The median 57 38.1% 42.5% cohort. included 864 108 Mean diameter 14.6mm 17.2mm In terms performance, AUC value ROC curves 96.0% 95.0% sensitivity model 97.6%, accuracy 97.5% specificity 97.6%. 90.2% 95.4%. Compared testing our had higher prediction (95.4% vs 88.9%). Conclusions: validated algorithm computer vision technique highly could assist rapid evaluation decisions.