Multimodal intraoperative diagnosis and grading system integrating macroscopic images, CT imaging, and textual reports for adenocarcinoma (MaCTex) in early-stage LUAD: A multicentric diagnostic study.

Abstract

8023 Background: Intraoperative frozen section (IFS) analysis is pivotal for guiding surgical strategies in stage IA lung adenocarcinoma (LUAD), specifically the extent of resection and lymph node dissection. However, IFS is constrained by sampling errors and prolonged turnaround times, which can compromise diagnostic precision and surgical efficiency. To address these unmet needs, we developed and validated a multimodal artificial intelligence framework. By integrating preoperative chest CT imaging, unstructured radiology text reports, and intraoperative macroscopic images of resected specimens, this system aims to deliver rapid, precise intraoperative predictions, thereby mitigating reliance on traditional IFS and optimizing surgical decision-making. Methods: This retrospective, multicenter study enrolled patients with stage IA LUAD who underwent complete resection between June 2020 and September 2023 across three institutions (Guangdong Provincial People's Hospital, Affiliated Hospital of Guangdong Medical University, and Meizhou People's Hospital). We collected preoperative thin-slice chest CT scans and unstructured text reports within three months prior to surgery. Intraoperative macroscopic images of resected specimens were captured via smartphone under natural lighting. We developed MaCTex, a multimodal AI model, to predict the IASLC grading system (PIL, MIA, IAC G1, G2, G3). Performance was evaluated against the gold standard of postoperative paraffin pathology. The study is registered with the Chinese Clinical Trial Registry (ChiCTR2500111776). Results: The cohort included 1,516 patients (yielding 1,638 pulmonary nodules) with matched preoperative CT imagings/reports and 2,344 intraoperative macroscopic images. We developed four distinct models for evaluation: a CT-only model, a CT-Text model, a Gross Image model, and the comprehensive MaCTex framework. The CT-Text model achieved a diagnostic AUC of 0.857, outperforming the unimodal CT model with an AUC of 0.837. The Gross Image model achieved a performance of 0.82 in five-class prediction. Notably, the integration of unstructured CT text reports significantly enhanced model efficacy, underscoring the critical value of radiologist expertise in refining intraoperative diagnostics. Conclusions: We established a robust multimodal AI framework integrating radiographic data, clinical text, and macroscopic pathology that enables efficient and accurate intraoperative prediction of IASLC grading in lung adenocarcinoma. By circumventing the sampling limitations of traditional methods, MaCTex serves as a promising alternative or adjunct to intraoperative frozen sections, providing thoracic surgeons with real-time, precise decision support to optimize surgical management.