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

Abstract

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.

Semi-supervised learning in cancer diagnostics

“…Several of the above-mentioned studies reported substantial performance gains for SSL as long as the model was short on labeled data, however, when the amount of labeled data was increased or only labeled data was used the gap between SSL and SL performance shrunk. However, the frequent lack of a comparison between baseline SL and SSL classifiers further complicates the evaluation of such studies and only few studies do report baseline comparisons ( 11 , 13 , 19 , 22 , 33 , 37 ) and still even fewer report equal tuning of hyperparameters ( 11 , 19 ) for SSL and SL classifiers to make results comparable. When it comes to model design, it is essential to note that different algorithms may perform differently with regard to different tasks ( 9 ).…”

“…Further, varying the amount of labeled and unlabeled data for both training and testing sets seems warranted to find the equilibrium of optimal performance for different tasks in future studies of SSL in oncology. The lack of reproducibility in research on artificial intelligence in general ( 44 ) is also likely to be a future issue in biomedical use-cases of SSL as unfortunately only a minority of studies provide publicly accessible code to support their results ( 11 , 19 , 27 , 28 , 30 , 38 , 40 ). As is evident from previous studies on SSL in oncology, use cases mainly include tumor entities with high prevalence such as breast ( 15 – 17 , 25 – 28 , 33 – 35 , 37 , 41 ), lung ( 18 , 22 , 23 , 33 , 34 , 38 , 41 ), and colorectal cancer ( 11 , 12 , 34 , 35 ) where single centers can amass sufficiently sized data sets to conduct SSL experiments.…”

“…The lack of reproducibility in research on artificial intelligence in general ( 44 ) is also likely to be a future issue in biomedical use-cases of SSL as unfortunately only a minority of studies provide publicly accessible code to support their results ( 11 , 19 , 27 , 28 , 30 , 38 , 40 ). As is evident from previous studies on SSL in oncology, use cases mainly include tumor entities with high prevalence such as breast ( 15 – 17 , 25 – 28 , 33 – 35 , 37 , 41 ), lung ( 18 , 22 , 23 , 33 , 34 , 38 , 41 ), and colorectal cancer ( 11 , 12 , 34 , 35 ) where single centers can amass sufficiently sized data sets to conduct SSL experiments. This is also reflected in the overwhelming absence of studies on SSL in hematology with only one single study ( 40 ) including any hematological neoplasm at all.…”

Liquid Biopsy and Artificial Intelligence as Tools to Detect Signatures of Colorectal Malignancies: A Modern Approach in Patient’s Stratification

“…Recently, Yu et al. ( 293 ) developed a recognition system for CRC which achieved one of the highest diagnosis accuracies in cancer diagnosis with AI using SL. More, AI has made a huge step in the field of intraoperative pathology, providing preliminary evaluations or highlighting suspicious areas ( 294 ).…”

Semi-supervised vision transformer with adaptive token sampling for breast cancer classification

“…Ma and Zhang (2018) developed an SSL model that combines affinity network fusion and a neural network to implement few-shot learning, significantly improving the model's learning ability with fewer training data. Other applications of SSL include cancer survival analysis (Liang et al, 2016), skin cancer diagnosis (Masood et al, 2015), bladder cancer grading (Wenger et al, 2022), and colorectal cancer detection (Yu et al, 2021). To our best knowledge, prior studies have not explored CT for BC detection, and our research aims to fill this gap.…”