RESUMO
Background: Lymph node metastasis (LNM) assessment in patients with papillary thyroid carcinoma (PTC) is of great value. This study aimed to develop a deep learning model applied to intraoperative frozen section for prediction of LNM in PTC patients. Methods: We established a deep-learning model (ThyNet-LNM) with the multiple-instance learning framework to predict LNM using whole slide images (WSIs) from intraoperative frozen sections of PTC. Data for the development and validation of ThyNet-LNM were retrospectively derived from four hospitals from January 2018 to December 2021. The ThyNet-LNM was trained using 1987 WSIs from 1120 patients obtained at the First Affiliated Hospital of Sun Yat-sen University. The ThyNet-LNM was then validated in the independent internal test set (479 WSIs from 280 patients) as well as three external test sets (1335 WSIs from 692 patients). The performance of ThyNet-LNM was further compared with preoperative ultrasound and computed tomography (CT). Findings: The area under the receiver operating characteristic curves (AUCs) of ThyNet-LNM were 0.80 (95% CI 0.74-0.84), 0.81 (95% CI 0.77-0.86), 0.76 (95% CI 0.68-0.83), and 0.81 (95% CI 0.75-0.85) in internal test set and three external test sets, respectively. The AUCs of ThyNet-LNM were significantly higher than those of ultrasound and CT or their combination in all four test sets (all P < 0.01). Of 397 clinically node-negative (cN0) patients, the rate of unnecessary lymph node dissection decreased from 56.4% to 14.9% by ThyNet-LNM. Interpretation: The ThyNet-LNM showed promising efficacy as a potential novel method in evaluating intraoperative LNM status, providing real-time guidance for decision. Furthermore, this led to a reduction of unnecessary lymph node dissection in cN0 patients. Funding: National Natural Science Foundation of China, Guangzhou Science and Technology Project, and Guangxi Medical High-level Key Talents Training "139" Program.
RESUMO
BACKGROUND: Microvascular invasion (MVI) is essential for the management of hepatocellular carcinoma (HCC). However, MVI is hard to evaluate in patients without sufficient peri-tumoral tissue samples, which account for over a half of HCC patients. METHODS: We established an MVI deep-learning (MVI-DL) model with a weakly supervised multiple-instance learning framework, to evaluate MVI status using only tumor tissues from the histological whole slide images (WSIs). A total of 350 HCC patients (2917 WSIs) from the First Affiliated Hospital of Sun Yat-sen University (FAHSYSU cohort) were divided into a training and test set. One hundred and twenty patients (504 WSIs) from Dongguan People's Hospital and Shunde Hospital of Southern Medical University (DG-SD cohort) formed an external test set. Unsupervised clustering and class activation mapping were applied to visualize the key histological features. RESULTS: In the FAHSYSU and DG-SD test set, the MVI-DL model achieved an AUC of 0.904 (95% CI 0.888-0.920) and 0.871 (95% CI 0.837-0.905), respectively. Visualization results showed that macrotrabecular architecture with rich blood sinus, rich tumor stroma and high intratumor heterogeneity were identified as the key features associated with MVI ( +), whereas severe immune infiltration and highly differentiated tumor cells were associated with MVI (-). In the simulation of patients with only one WSI or biopsies only, the AUC of the MVI-DL model reached 0.875 (95% CI 0.855-0.895) and 0.879 (95% CI 0.853-0.906), respectively. CONCLUSION: The effective, interpretable MVI-DL model has potential as an important tool with practical clinical applicability in evaluating MVI status from the tumor areas on the histological slides.