The development of skin lesion detection application in smart handheld devices using deep neural networks

Early detection of malignant skin lesions improves patient survival rates. Conventional self-detection method for public invariably suffers from limitations: subjectivity, inaccuracy, and expert dependent variability. Therefore, this study presents a detailed development workflow to establish a multimedia-based healthcare systems using computational intelligence, specifically, a mobile application with skin lesion detection capability by integrating state-of-the-art deep learning frameworks that facilitates the global users to execute malignant skin lesions self-detection using a smartphone. We applied transfer learning on various object detection models using ISIC skin lesions dataset with TensorFlow Object Detection API. The selected object detection model is SSD MobileNetV2 with 93.9% of evaluation accuracy. The trained object detection model has been successfully integrated into the mobile application using Firebase ML Kit and has reported low detection time on smartphones. The mobile application has tested to be compatible with various Android versions and screen sizes after we experimented with Firebase Test Lab using seven different smartphones. The trained deep learning model and mobile application development project can be obtained from Github ( https://github.com/UTARSL1/ ).