Medmain Co., Ltd.
Succeeded in developing artificial intelligence that enables discrimination of malignant melanoma in digital skin histopathological specimens ~ Paper published in Cancers ~
Medmain Co., Ltd. (Headquarters: Fukuoka City, Fukuoka Prefecture, CEO: Osamu Iizuka, hereinafter referred to as “Medmain”), which provides the digital pathology support solution “PidPort”, uses deep learning to We have succeeded in developing artificial intelligence that enables the differentiation of malignant melanoma in
histopathological digital specimens.
The key point of this research result is that, by using deep learning, it has become possible to differentiate malignant melanoma of the skin, a rare cancer with an overwhelmingly small number of cases in Japan. In addition, although the differentiation between malignant melanoma such as Spitz’s nevus and diseases that require
differentiation is insufficient, by performing tile-level evaluation, those that can be distinguished from malignant melanoma can be precisely malignant melanoma. It is now possible to evaluate the infiltration range of tumor cells.
We are pleased to inform you that this research has been published in Cancers published by MDPI (https://www.mdpi.com).
(URL of paper: https://www.mdpi.com/2072-6694/15/6/1907)
[Image 1: https://prtimes.jp/i/34505/33/resize/d34505-33-664c6062e2db1ced951e-2.png&s3=34505-33-b7867d06e7a3fcae43c64e2cc87098da-1005×636.png] ■Summary of this research result We have succeeded in developing an artificial intelligence that can differentiate malignant melanoma from skin pathological tissue digital specimens. ■Background of this research Malignant melanoma that develops on the skin is a malignant tumor (cancer) derived from melanocytes, and is considered to be one of the malignant tumors with the worst prognosis. Most malignant melanomas of the skin appear on the surface of the skin as brown to black pigmented spots (blemishes) or masses. In general, “asymmetric shape”, “unclear outline”, “uneven color”, “6 mm or more in size”, “larger”, “change in color, shape, hardness, etc.” characteristics can be seen. There are racial differences in onset, and while the frequency of onset is high in Westerners, it is treated as a “rare cancer” because the incidence is low, 1 to 2 per 100,000 Japanese. Malignant melanoma can often be distinguished from benign nevus (mole) or pigmented spots by macroscopic changes, and dermoscopy can be used to make a more accurate diagnosis. On the other hand, if the lesion cannot be diagnosed by macroscopic examination, all or part of the lesion may be surgically removed and histopathological examination may be performed. In histopathological examination, it is necessary to clarify the difference from lesions that require differential diagnosis such as benign nevus. It is Because the number of “rare cancers” including malignant melanoma is small, there are far more problems in diagnosis and treatment compared to other cancers, and the development of innovative diagnostic and treatment methods is desired. increase. Based on the above clinical background, in this research, we decided to develop artificial intelligence that can distinguish malignant melanoma from digital skin histopathological specimens using deep learning. ■Details of this research We digitized skin
pathological tissue specimens (WSI: Whole-Slide Image) provided by domestic facilities and created supervised data including annotation data by multiple pathologists. The number of WSI data used for this development was 78 for cutaneous malignant melanoma and 88 for lesions other than malignant melanoma (nevus, benign lesions other than nevus, etc.). Since malignant melanoma is a rare cancer in Japan, the number of cases that can be collected is limited. . 38 out of 78 cutaneous malignant melanomas were used as training data, and 38 out of 88 lesions other than melanoma were used as training data. In this AI development, in addition to the previous development, we also performed evaluation at a minute tile level to evaluate the
infiltration range of malignant melanoma. Results of this study When fully supervised learning using pathologist’s annotation data was added to weakly supervised learning, an ROC-AUC value of 0.825 was obtained at the WSI level (evaluation per slide glass). was given. In addition, a high ROC-AUC value of 0.936 was obtained when evaluation was performed at the divided tile level to evaluate the infiltration range of malignant melanoma. In addition, the areas of malignant melanoma identified by artificial intelligence displayed by the heatmap were validated by multiple pathologists. On the other hand, some difficult-to-distinguish nevus such as Spitz nevus and blue nevus are false-positive, and melanoma in-situ, which is extremely difficult to diagnose, is false-negative. issues remained.
[Image 2: https://prtimes.jp/i/34505/33/resize/d34505-33-a458509280039c620cf0-2.jpg&s3=34505-33-6f7d1a0b87b08136cba354c8a475711d-2362×1159.jpg] ■Future Prospects We will proceed with verification of the deep learning artificial intelligence model developed this time at multiple facilities and large-scale cases. The false positives and false negatives that occurred in this research are expected to be improved by increasing the number of cases and conducting additional learning, so we will continue to advance research and development. Through this research, we have found a way to develop artificial intelligence even for rare cancers, which have a small number of cases. We will continue to contribute to ■Original paper ▼Paper title: Deep Learning Approach to Classify Cutaneous Melanoma in a Whole Slide Image ▼Japanese translation: Development of artificial intelligence using deep learning to differentiate malignant melanoma in skin histopathological digital specimens ▼DOI: https: //www.mdpi.com/2072-6694/15/6/1907 Authors and affiliations -Department of Pathology, Tochigi Cancer Center- Makoto Abe -Department of Pathology, Tokyo Shinagawa Hospital- Morio Nakano -Medmain Co., Ltd.- Jo Masayuki Ki, Meng Li *This result was obtained as a result of a subsidized project by the New Energy and Industrial Technology Development Organization (NEDO), a national research and development agency. ■Company Profile [Company name] Medmain Inc. *Ministry of Economy, Trade and Industry J-START UP selected company https://www.j-startup.go.jp/startups [Establishment date] January 11, 2018 [Business] Planning, development, operation and sales of medical software and cloud services [Representative Director/CEO] Osamu Iizuka [Location] [Tokyo Office] 2-10-11 Minami-Aoyama, Minato-ku, Tokyo Aoyama Building 2F / [Fukuoka] Office] 2-4-5 Akasaka, Chuo-ku, Fukuoka City, Fukuoka Prefecture Chatelet Succeeds 104 ■ Various related sites [Corporate site]
https://medmain.com [Product site] https://service.medmain.com ■ Contact information Medmain Co., Ltd. Public Relations:
pr-m@medmain.com
Details about this release:
https://prtimes.jp/main/html/rd/p/000000033.000034505.html
MAIL:cr@prtimes.co.jp