Adaptively learning visual model of surgical resection images via coarse-to-fine–grained strategy for pathological invasiveness prediction in indeterminate stage IA lung adenocarcinoma.

Abstract

1563 Background: Indeterminate nodules with high risk of malignancy are suggested to receive video-assisted thoracoscopy biopsy. Frozen section assists surgeons to decide surgical strategies for stage IA lung adenocarcinoma (LUAD) intraoperatively, which is time-consuming and may misdiagnosis due to limited sampling and poor tissue quality. Methods: We prospectively enrolled stage IA LUAD patients underwent complete resection between June 2020 and September 2023 in Guangdong Provincial People's Hospital (GDPH), Affiliated Hospital of Guangdong Medical University, Meizhou People's Hospital. All the resected lung nodules were finally diagnosed as pre-invasive lesion (PIL), minimally invasive adenocarcinoma (MIA) or invasive lung adenocarcinoma (IAC) by FFPE pathological diagnosis. Images of the whole section of the lung nodules were taken by smartphones under natural lighting condition in operating theaters without shadowless surgical lights. Patients received preoperative anti-cancer therapy and with low-quality tumor section images were excluded. Images of nodules section, FS reports, FFPE diagnosis and clinical information were prospectively collected. Predictive artificial intelligence (AI) model was developed by a three-step process, which first predict the IAC and non-IAC region at coarse level, and identify fine level features through the high-risk region. The coarse and fine labels were then coordinated through risk ranking loss to reach a final predictive result. Results: We prospectively enrolled 1516 patients with 1638 indetermined lung nodules from preoperative chest CT following the inclusion and exclusion criteria, with 2438 images acquired from intraoperative nodule sections. The included patients have median age of 57 years old and 38.6% male patients. The pathological diagnosis for included lung nodules are 122 PIL nodules, 502 MIA nodules and 1014 stage IA LUAD. A multi-task artificial intelligence based on Coarse-to-Fine-Grained Strategy has been established to classify the indetermined lung nodules during the surgery. For binary classification, AI model reach an AUC of 0.86 in differentiation between IAC or non-IAC. For 3-classification level, AI model reach an AUC of 0.87 in differentiation between PIL, MIA or IAC. We further defined the IAC nodules into G1, G2 and G3 according to IASLC lung adenocarcinoma grading system. For 5-classification level, AI model reach an AUC of 0.86 in differentiation between PIL, MIA, IAC-G1, IAC-G2 and IAC-G3. Conclusions: Our AI models based on images of intraoperative surgical resection could effectively classify indetermined lung nodules, which can assist thoracic surgeons diagnose the nodules rapidly during the surging and decide the following surgical strategies. Clinical trial information: ChiCTR2300075999 .