Deep learning model for pathological invasiveness prediction using smartphone-based surgical resection images in
Lintong Yao1, Linghan Cai2, Maotao Weng1
1Department of Thoracic Surgery, Guangdong Provincial People's Hospital (Guangdong Academy of Medical Sciences), Southern Medical University, Guangzhou, China.
Background:
In clinical stage IA lung adenocarcinoma (LUAD), rapid and accurate intraoperative diagnosis is crucial to decide whether to perform segmentectomy and lobectomy. Frozen section analysis is time consuming and not always reliable for LUAD diagnosis and grading. We developed deep learning models using surgical resection images to assist in prompt diagnosis and risk stratification of stage IA LUAD to aid in surgical decision making.
Methods:
In this prospective, multicentre cohort, patients with clinical stage IA LUAD were enrolled from June 1, 2020, to Sept 30, 2023, from three hospitals in China. Surgical resection images of LUAD were captured using smartphones under natural lighting conditions in the operating theatre. Deep learning models were established based on these images for three tasks: identification of invasive lung adenocarcinoma from non-invasive lung adenocarcinoma lesions; diagnosis of adenocarcinoma in situ, minimally invasive adenocarcinoma, and invasive lung adenocarcinoma; and grading of adenocarcinoma in situ, minimally invasive adenocarcinoma, invasive lung adenocarcinoma grade 1, grade 2, and grade 3 according to the International Association for the Study of Lung Cancer Grading System. The study is registered with the Chinese Clinical Trial Registry, ChiCTR2300075999.
Findings:
We enrolled 1529 patients with 2344 surgical section images from Guangdong Provincial People's Hospital, 116 patients with 307 images from Affiliated Hospital of Guangdong Medical University, and 82 patients with 259 images from Meizhou People's Hospital. The area under curve for the surgical resection image-based model (SuRImage) was 0·84 (95% CI 0·82-0·86) for invasive lung adenocarcinoma identification, 0·87 (0·86-0·88) for invasive lung adenocarcinoma diagnosis, and 0·85 (0·83-0·86) for invasive lung adenocarcinoma grading in Guangdong Provincial People's Hospital. SuRImage showed better diagnosis performance than frozen section. Assisted with SuRImage, average diagnostic accuracy of thoracic surgeons could be improved from 63·80% (95% CI 60·57-67·03) to 73·44% (67·68-79·19) for invasive lung adenocarcinoma grading.
Interpretation:
This first-in-field diagnostic study focused on intraoperative diagnosis based on surgical resection images in stage IA LUAD and provides insights into the macroscopic morphological features for pathological invasiveness. By elucidating macroscopic morphological indicators of invasiveness, SuRImage empowers surgeons to make more precise, timely decisions, optimise intervention strategies, and streamline the surgical workflow.
Funding:
National Key R&D Program of China; National Natural Science Foundation of China; International Science and Technology Cooperation Program of Guangdong; Natural Science Foundation of Guangdong; Beijing Xisike Clinical Oncology Research Foundation; Meizhou Medical and Health Scientific Research Projects.