Diagnostic Assessment Of Deep Learning Algorithm For Melanocytic Tumors Using Pathological Images

Background:Melanoma is one of the most common malignant tumors that originated from melanocytes.The incidence of melanoma among Chinese continues to be increasing.Accurate melanoma diagnosis is an essential task performed by pathologists worldwide to inform clinical management.This process requires highly skilled pathologists and is fairly time consuming and error prone.However,the hospitals in China generally face a situation of serious shortage of pathologists.And it takes a long time to cultivate an experieneed pathologist.Rapid and accurate inte:r:pretation of slides is not always available,particularly in the remote areas.The development of the deep learning algorithm has allowed for significant gains in the ability to classify images and detect objects in a picture.Application of deep learning to whole slide pathological images(WS5s)can potentially improve diagnostic accuracy and efficiency.Objective:To assess the performance of deep learning algorithm for the classification of melanoma and nevus using WSIs.Methods:A total of 753 melanocytic skin lesions(114 melanomas,639 nevi)were retrospectively sampled from patients who underwent surgery for clinical care at three skin pathology centers in China.The dataset represented the most conmon melanocytic tumors encountered in routine pathology practice.Three experienced dermatopathologists(each with over 20 years of cutaneous pathology experience)independently reviewed the glass slides on their own microscopes while blinded to others' interpretations and entered their diagnoses into one of five Melanoma Pathology Assessment&Treatment Hierarchy(MPATH-Dx)categories.The dermatopathologists then discussed the cases,using a modified Delphi method to facilitated consensus building for cases with discordant diagnoses.Stratified random sampling was performed on the basis of the 5 categories according to MPATH-Dx of the 753 WSIs.A total of 649 WSIs were used to train the algorithm.Another 104 WSIs were used to test the screening performance of the algorithm.The DenseNetl 69 architecture was adopted as the classification algorithm,which was pretrained on approximately 1.28 million images in ImageNet and finetuned using the training dataset of 649 WSIs(85 melanomas,564 nevi).Diagnostic classification was validated in an independent test set of 104 WSIs(29 melanomas,75 nevi)by the algorithm and a panel of 7 pathologists.Results:The 7 pathologists interpreted these 104 WSIs independently without knowledge of clinical data and the others' diagnoses.The mean[±standard deviation(SD)]sensitivity and specificity of the 7 pathologists for the dichotomous classification were 85.2%(±3.8%)and 95.6%(±1.5%),respectively.This translated into an average(±SD)AUC(area under receiver operating characteristic curve)of 0.90(±0.22).Using the pathologists' mean sensitivity of 85.2%as the operating point on the algorithm's ROC(receiver operating characteristic)curve,there was no significant difference between the specificity of deep learning algorithm and the mean specificity of pathologists(96.0%vs.95.6%,P=0.52).Using the pathologists' mean specificity of 95.6%as the operating point on the algorithm's ROC curve,the deep learning algorithm's sensitivity was higher than the mean sensitivity of pathologists(93.1%vs.85.2%,P<0.01).Moreover,the algorithm's AUC was greater than the mean AUC of the pathologists(0.97 vs.0.90,P<0.01).We also provided a more transparent and interpretable diagnosis by highlighting the regions of interest recognized by the neural network in WSIs.The deep learning algorithm achieved the Jaccard of 77.78%,which means that most of the regions of interest were identified correctly by the algorithm.Conclusion:Deep learning algorithm might be as a supplemental screening tool to increase the pretest probability of a biopsy demonstrating melanoma.Whether this approach has clinical utility will require evaluation in a prospective clinical setting.