Skin Disease Classification with Point Guidance Training

Abstract

Melanoma, a type of skin cancer, claims over 9,000 lives every year. Many melanomas are initially detected by patients themselves, but distinguishing these from benign and common skin diseases like moles and seborrheic keratosis can be challenging for non-experts. Consequently, there is a pressing need to develop artificial intelligence capable of differentiating melanoma from similar skin diseases. However, this endeavor faces several challenges. Firstly, medical datasets, such as those containing images of skin diseases, are often small due to privacy concerns and the high cost of labeling. Furthermore, melanoma is relatively rare compared to conditions like moles, resulting in even more limited data within these datasets. To address these issues, we proposed a training method that utilizes point guidance alongside class labels to guide the model. On the ISIC 2017 skin disease dataset, our point-guided approach enhances the model’s ability to localize objects and improves overall performance, even when the training data is limited or imbalanced, compared to the baseline model. Additionally, we demonstrate that combining point guidance with active learning can significantly reduce the amount of training data required to achieve the desired performance. Our findings indicate that only 40% to 60% of the data is necessary to match the baseline’s performance. Through point-guided model training, we aim to contribute to the development of diagnostic models for melanoma, effectively reducing labeling costs.