AbstractPurposesThe machine-assisted recognition of colorectal cancer using pathological images has been mainly focused on supervised learning approaches that suffer from a significant bottleneck of requiring a large number of labeled training images. The process of generating high quality image labels is time-consuming, labor-intensive, and thus lags behind the quick accumulation of pathological images. We hypothesize that semi-supervised deep learning, a method that leverages a small number of labeled images together with a large quantity of unlabeled images, can provide a powerful alternative strategy for colorectal cancer recognition.MethodWe proposed semi-supervised classifiers based on deep learning that provide pathological predictions at both patch-level and the level of whole slide image (WSI). First, we developed a semi-supervised deep learning framework based on the mean teacher method, to predict the cancer probability of an individual patch by utilizing patch-level data generated by dividing a WSI into many patches. Second, we developed a patient-level method utilizing a cluster-based and positive sensitivity strategy on WSIs to predict whether the WSI or the associated patient has cancer or not. We demonstrated the general utility of the semi-supervised learning method for colorectal cancer prediction utilizing a large data set (13,111 WSIs from 8,803 subjects) gathered from 13 centers across China, the United States and Germany. On this data set, we compared the performances of our proposed semi-supervised learning method with those from the prevailing supervised learning methods and six professional pathologists.ResultsOur results confirmed that semi-supervised learning model overperformed supervised learning models when a small portion of massive data was labeled, and performed as well as a supervised learning model when using massive labeled data. Specifically, when a small amount of training patches (~3,150) was labeled, the proposed semi-supervised learning model plus ~40,950 unlabeled patches performed better than the supervised learning model (AUC: 0.90 ± 0.06 vs. 0.84 ± 0.07,P value = 0.02). When more labeled training patches (~6,300) were available, the semi-supervised learning model plus ~37,800 unlabeled patches still performed significantly better than a supervised learning model (AUC: 0.98 ± 0.01vs. 0.92 ± 0.04, P value = 0.0004), and its performance had no significant difference compared with a supervised learning model trained on massive labeled patches (~44,100) (AUC: 0.98 ± 0.01 vs. 0.987 ± 0.01, P value = 0.134). Through extensive patient-level testing of 12,183 WSIs in 12 centers, we found no significant difference on patient-level diagnoses between the semi-supervised learning model (~6,300 labeled, ~37,800 unlabeled training patches) and a supervised learning model (~44,100 labeled training patches) (average AUC: 97.40% vs. 97.96%, P value = 0.117). Moreover, the diagnosis accuracy of the semi-supervised learning model was close to that of human pathologists (average AUC: 97.17% vs. 96.91%).ConclusionsWe reported that semi-supervised learning can achieve excellent performance at patch-level and patient-level diagnoses for colorectal cancer through a multi-center study. This finding is particularly useful since massive labeled data are usually not readily available. We demonstrated that our newly proposed semi-supervised learning method can accurately predict colorectal cancer that matched the average accuracy of pathologists. We thus suggested that semi-supervised learning has great potentials to build artificial intelligence (AI) platforms for medical sciences and clinical practices including pathological diagnosis. These new platforms will dramatically reduce the cost and the number of labeled data required for training, which in turn will allow for broader adoptions of AI-empowered systems for cancer image analyses.
Reading Preferences of ESL Students: Electronic Texts vs. Printed
