Accurate recognition of colorectal cancer with semi-supervised deep learning on pathological images

• Published in: Nature communications Vol. 12; no. 1; pp. 6311 - 13
• Main Authors: Yu, Gang, Sun, Kai, Xu, Chao, Shi, Xing-Hua, Wu, Chong, Xie, Ting, Meng, Run-Qi, Meng, Xiang-He, Wang, Kuan-Song, Xiao, Hong-Mei, Deng, Hong-Wen
• Format: Journal Article
• Language: English
• Published: London Nature Publishing Group UK 02.11.2021; Nature Publishing Group; Nature Portfolio
• Subjects: 119/118; 631/114/1305; 631/67/2321; 639/166/985; Annotations; Artificial intelligence; Artificial Intelligence - standards; Cancer; Colorectal cancer; Colorectal carcinoma; Colorectal Neoplasms - classification; Colorectal Neoplasms - diagnostic imaging; Colorectal Neoplasms - pathology; Deep learning; Deep Learning - standards; Humanities and Social Sciences; Humans; Learning algorithms; Lung Neoplasms - classification; Lung Neoplasms - diagnostic imaging; Lung Neoplasms - pathology; Lymph nodes; Lymphatic Metastasis; multidisciplinary; Neural Networks, Computer; Object recognition; ROC Curve; Science; Science (multidisciplinary); Semi-supervised learning; Supervised Machine Learning - standards; Teachers
• ISSN: 2041-1723; 2041-1723
• DOI: 10.1038/s41467-021-26643-8

Machine-assisted pathological recognition has been focused on supervised learning (SL) that suffers from a significant annotation bottleneck. We propose a semi-supervised learning (SSL) method based on the mean teacher architecture using 13,111 whole slide images of colorectal cancer from 8803 subjects from 13 independent centers. SSL (~3150 labeled, ~40,950 unlabeled; ~6300 labeled, ~37,800 unlabeled patches) performs significantly better than the SL. No significant difference is found between SSL (~6300 labeled, ~37,800 unlabeled) and SL (~44,100 labeled) at patch-level diagnoses (area under the curve (AUC): 0.980 ± 0.014 vs. 0.987 ± 0.008, P value = 0.134) and patient-level diagnoses (AUC: 0.974 ± 0.013 vs. 0.980 ± 0.010, P value = 0.117), which is close to human pathologists (average AUC: 0.969). The evaluation on 15,000 lung and 294,912 lymph node images also confirm SSL can achieve similar performance as that of SL with massive annotations. SSL dramatically reduces the annotations, which has great potential to effectively build expert-level pathological artificial intelligence platforms in practice. Machine-assisted recognition of colorectal cancer has been mainly focused on supervised deep learning that suffers from a significant bottleneck of requiring massive amounts of labeled data. Here, the authors propose a semi-supervised model based on the mean teacher architecture that provides pathological predictions at both patch- and patient-levels.