Improving Teledermatology Referral with Edge-AI: Mobile App to Foster Skin Lesion Imaging Standardization

Every year, the number of skin cancer cases has been increasing which, consequently, increases the strain on the health care systems around the globe. With the growth of processing power and camera quality on smartphones, the investment in telemedicine and the development of mobile teledermatology applications can, not only contribute to the standardization of image acquisitions but also, facilitate early diagnosis. This paper presents a new process for real-time automated image acquisition of macroscopic skin images with the merging of an automated focus assessment feature-based machine learning algorithm with conventional computer vision techniques to segment dermatological images. Three datasets were used to develop and evaluate the proposed methodology. One comprised of 3428 images acquired with a mobile phone for this purpose and 1380 from the other two datasets which are publicly available. The best focus assessment model achieved an accuracy of 88.3% and an F1-Score of 86.8%. The segmentation algorithm obtained a Jaccard index of 85.81% for the SMARTSKINS dataset and 68.59% for the Dermofit dataset. The algorithms were deployed to a mobile application, available in Android and iOS, without causing any performance hindrances. The application was tested in a real environment, being used in a 10-month pilot study with six General and Family Medicine doctors and one Dermatologist. The easiness to acquire dermatological images, image quality, and standardization were referred to as the main advantages of the application.

Automated summary

This paper presents a new method for real-time automated image acquisition and segmentation of macroscopic skin images using a combination of feature-based machine learning algorithm and conventional computer vision techniques. The proposed methodology was developed and evaluated using mobile phone images and publicly available datasets, achieving high accuracy and effectiveness. The algorithms were successfully implemented into a mobile application and tested in a real environment with positive feedback.