02726nas a2200289   4500000000100000008004100001260003900042653003600081653000700117653002200124653002200146653002000168653001700188100001500205700001500220700001300235700001400248700001200262700001700274700001500291700002400306245010200330856009000432300001100522490000700533520189600540       2023                            d  bHind Kusht Nivaran SangaNew Delhi10aApplied Artificial Intelligence10aAI10aFew Shot Learning10aLeprosy Screening10aSiamese Network10aSkin Imaging1 aBeesetty R1 aReddy S. A1 aModali S1 aSunkara G1 aDalal J1 aDamagathla J1 aBanerjee D1 aVenkatachalapathy M00aLeprosy Skin Lesion Detection: An AI Approach Using Few Shot Learning in a Small Clinical Dataset  uhttps://www.ijl.org.in/published-articles/29062023223603/1-R-Beesetty-et-al-Final.pdf  a89-1020 v953 a<p>This is an exploratory research study to check if artificial intelligence (AI) based image marker tool can aid leprosy screening to detect leprosy cases early in field situation and reduce the financial and personnel burden. We aimed to collect clinical leprosy skin lesion images and develop an AI model to identify and differentiate them. A total of 368 clinically diagnosed leprosy and 28 non-leprosy skin lesions were collected by an expert leprologist from 151 eligible patients using a multimodal imaging protocol. A Siamese-based Few Shot Learning (FSL) model was trained as it is a meta learning approach on an extremely small data set with fewer disease classes (disease conditions as categories). The number of class labels were increased by fine-grained grouping of skin lesions based on skin morphology (Nine leprosy subgroups) and further divided into train-set and test-set. An AI model was successfully developed, and the results indicated an accuracy of 91.25% and 73.12% on train-set and test-set for two-way one-shot task, respectively. The best sensitivity-specificity for the test-set were 72.39%-73.66% (two-way one-shot task). This early research data indicates that the development of AI based leprosy screening application is feasible using the skin lesion image as marker. The FSL method was successfully used in this training the small data set. However, this is a small sample size study, and more leprosy cases need to be enrolled along with an equal number of non-leprosy cases while improving model architecture to reduce overfit or bias problem. Moreover, as of now this tool cannot be used for neural leprosy (having no skin lesion) as well as lepromatous leprosy having diffuse infiltration. This tool will need further development and validation on pictures taken by different categories of common health care workers using available mobile phones.</p>