297 Deep Ackerman a novel deep learning method to develop dermatopathology diagnosis by artificial intelligence

Abstract

Recent progress of deep convolutional neural networks (CNNs) proposed by J. Hinton in 2012 succeeded in classifying general images and opened the possibility for automated diagnosis of medical images. However, the application of CNNs to histopathology image classification is, a challenging task. Since the pixel size of virtual slide images, which microscopically scanned a whole histopathology slide, are comprised of 40 to 100 mega pixels and too large to input into usual CNN. To overcome the issue of large pixel files, we developed a novel method, which consists of two CNN models. First, a whole histopathology slide image is dissected into micro images containing one cell per image. Then, the first CNN classifies dissected images into cell and structure types in four categories of atypical cells, epidermis, dermis, and subcutaneous tissue. Based on the result of the first CNN, heatmaps of each cell or structure class are constructed. Then, the second CNN model diagnosis based on the heatmaps, like as Dr. Bernard Ackermans histopathology illustration. We used 60,000 divided micro images of one malignant melanoma virtual slide image as a dataset for the first CNN deep learning. We convert 121 slides of malignant melanoma and 21 slides of nevus pigmentosus to heatmaps and inputted them to the second CNN. Our model showed 73 % of F-score, while a naïve CNN approach that directly input whole histopathology slide images into single CNN achieved 57 % of F-score. We also examined if our method can diagnose other skin tumors, such as basal cell carcinoma, squamous cell carcinoma, seborrheic keratosis, epidermal cyst, Bowens disease, extramammary Pagets disease, and mycosis fungoides. Our result indicated that dermatopathology diagnosis by artificial intelligence requires strategical approaches to reach the levels of experienced dermatopathologists.

Image analysis of cutaneous melanoma histology: a systematic review and meta-analysis

“…Overall, 4,888 specimens were included, of which at least 2,715 were melanoma specimens. The diagnostic entities within the datasets varied between studies, with some only containing melanoma deposits 12,21,[24][25][26]33 and others containing more than one pathology 10,11,22,[27][28][29][30][31][32] .…”

“…Most studies focused on a binary classification task, with some focussing on detection and localisation of melanoma deposits in WSIs containing melanoma (melanoma versus not melanoma) 12,25,26,32 and others performing diagnostic classifications including melanomas versus naevi 11,22,31 and primary melanoma versus metastatic melanoma 33 . Five studies addressed more complex classifications into three or more diagnostics entities 10,23,[28][29][30] . One study 24 did not focus on a classification task, but instead studied automation of the proliferation index in melanoma.…”

“…There was some degree of variation in the IT used. Most employed the use of a convolutional neural networks (CNN), with the architecture differing considerably [10][11][12]22,23,26,[29][30][31]33 . Studies using CNNs were more recently conducted.…”

“…Two studies used a combination of a CNN and SVM as their index test 11,32 . A further study used more basic image processing and adaptive thresholding method 29 .…”

“…Some studies used a combination of these approaches 23,32 . Two studies did not detail the reference standard used 29,33 .…”