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

Abstract

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