Pradhumn Agrahari

Skin cancer is a form of cancer that is a major health concern with 125,000 new melanoma cases every year. Even around 3 million non-melanoma cases can be seen worldwide each year. Initially, diagnosis of skin lesion was done visually by employing dermoscopic analysis, clinical screening, and several other methods. It is seen that inexperienced dermatologists can subsequently reduce the diagnostic accuracy of skin lesions. Early detection of skin cancer can decrease the mortality rate. Earlier… Show more

Skin cancer is a form of cancer that is a major health concern with 125,000 new melanoma cases every year. Even around 3 million non-melanoma cases can be seen worldwide each year. Initially, diagnosis of skin lesion was done visually by employing dermoscopic analysis, clinical screening, and several other methods. It is seen that inexperienced dermatologists can subsequently reduce the diagnostic accuracy of skin lesions. Early detection of skin cancer can decrease the mortality rate. Earlier studies have demonstrated that deep learning has given a better performance than human experts in various computer vision tasks. The paper proposes an effective multiclass skin cancer detection system having high performance comparable to that of a dermatology expert. A pre-trained MobileNet model is employed for model building. For training HAM10000 ISIC dataset is used. The model detects skin lesion with a categorical accuracy as high as 80.81%, top-2 accuracy of 91.25%, and top-3 accuracy of 96.26%. This is a fast and expansible method that can hold the ability for impeccable clinical advancement, including widening the scope and scale of primary healthcare practice. Show less