Automated differentiation of skin melanocytes from keratinocytes in high‐resolution histopathology images using a weakly‐supervised deep‐learning framework. Issue 1 (24th September 2022)

• Record Type: Journal Article
• Authors: Parajuli, Madan; Shaban, Mohamed; Phung, Thuy L.
• Is Part Of: International journal of imaging systems and technology. Volume 33:Issue 1(2023)
• Page Start: 262
• Page End: 275
• Publication Date: 2022-09-24
• Subjects: artificial intelligence -- deep‐machine learning -- digital pathology -- keratinocytes -- melanocytes; Imaging systems -- Periodicals; Image processing -- Periodicals; 621.367
• DOI: 10.1002/ima.22810
• Languages: English
• ISSNs: 0899-9457

Abstract

Abstract: Early detection and accurate diagnosis of melanoma skin cancer which accounts for 75% of all skin cancer mortalities would significantly improve survival rates. Melanoma is characterized by an abnormal proliferation of malignant melanocytes in the epidermis of the skin, which is composed predominantly of keratinocytes. Localization and differentiation of melanocytes from keratinocytes in the skin in whole slide images (WSIs) is an important initial task towards the diagnosis of the tumor. Because of the small size of melanocytes and the dominance of keratinocytes in the epidermis, identification of melanocytes can be challenging and prone to errors. We propose a new fully‐automated framework based on deep‐machine learning to identify melanocytes and keratinocytes with detection accuracy of 90.5% and 87.4%, respectively. This framework begins with segmenting the epidermis layer inside 640 × 320 × 3 super tiles of WSIs using a DeepLabV3+ model followed by a weakly‐supervised deep learning approach deploying a fine‐tuned pre‐trained visual geometry group (VGG)‐16 model, gradient‐weighted class activation mapping (Grad‐CAM), Otsu's and contour estimation methods in order to identify melanocytes and keratinocytes inside each 64 × 64 × 3 tile (i.e., a subdivision of the super tile). The proposed framework outperforms the state‐of‐the‐art methods as well as provides a significant improvement of 28% and 17% over the fully supervised faster region‐based convolutional neural network (R‐CNN) in detecting melanocytes and keratinocytes respectively without the need for expensive expert fine labels for model training and validation. The proposed framework offers a promising accurate tool to aid pathologists in differentiating melanocytes from keratinocytes that would eventually support the diagnosis of melanoma.