Abstract

Millions of people across the globe contract skin conditions and rashes that may irritate, clog, or inflame skin, subsequently leading to redness, swelling, burning, and itching if not treated immediately. In developing countries, there is a great need for an automated diagnosis system that would reduce manual efforts and the time consumption of dermatologists and patients. A computer-automated detection of such skin diseases is possible by taking a computer vision approach. The ensemble deep learning pipeline utilizes a 34- layer ResNet, a popular image classification convolutional neural network, to detect 11 major skin conditions with only pixels and labels as inputs. With no previously made training datasets available for the research of skin diseases, the dataset for the model was designed from scratch and involved gathering 587 total images of 11 classes of skin rashes and conditions, while also accounting for healthy skin in a miscellaneous class. The calculated average precision for my ResNet model was 0.917 with precision (strength against false positives) as 0.925 and with recall (strength against false negatives) being 0.681. My implementation of an adjusted, optimized ResNet model proved successful in that it was possible to beat state-of-the-art online architectures such as Google AutoML’s API which had a slightly lower average precision of 0.887.

Full Text
Published version (Free)

Talk to us

Join us for a 30 min session where you can share your feedback and ask us any queries you have

Schedule a call