Intraoperative frozen section pathological diagnosis of lung adenocarcinoma serves as the gold standard for determining the extent of surgical resection. Due to the dual constraints of intraoperative time limitations and the challenge of manually assessing tumor invasion volume in three-dimensional space, current manual diagnostic approaches and weakly supervised deep learning methods have demonstrated suboptimal diagnostic accuracy. To enhance the accuracy of intraoperative pathological diagnosis of lung adenocarcinoma and provide more precise recommendations for intraoperative resection extent to thoracic surgeons, we have developed a Hybrid-Supervised Framework for Lung Adenocarcinoma (HSFLA). This framework accomplishes the following processes: hybrid-supervised diagnosis of 2D whole slide images (WSIs), automatic annotation of tumor invasive regions, registration of consecutive WSIs, and 3D reconstruction and volume calculation of tumor invasive areas. We evaluated HSFLA on a dataset comprising 1161 WSIs from two centers and three subtypes, achieving an accuracy of 95.6%, which represents a significant improvement over manual review (84.7%) and weakly supervised learning (66.2% ± 3.0%). The consistency of its invasive area automatic annotations with manual pixel annotations was 86.6%. Furthermore, HSFLA’s concordance with spatial transcriptomics samples demonstrated its interpretability at the genetic level. Utilizing HSFLA’s automatic annotation functionality provided pathologists with a safe and effective diagnostic aid, improving their manual diagnostic accuracy by 22.9% (n = 3). We also applied HSFLA in real-world clinical settings for prospective study. Compared to manual diagnosis alone, the “human-machine interaction” diagnostic mode provided more appropriate surgical recommendations for 5 patients (among 70). Overall, HSFLA demonstrates the potential clinical utility of artificial intelligence in supporting intraoperative pathological assessment and surgical decision-making, and may serve as a paradigm for future innovations in AI-assisted clinical workflows.
Zhao et al. (Mon,) studied this question.