An Embedded Vision System for Autoimmune Skin Diseases Classification Based on Deep Learning: A Preliminary Study

The skin can exhibit symptoms of internal patho-logic conditions. Cutaneous autoimmunity is one such instance, serving as a sign of a systemic lack of immune tolerance. It is quite challenging to appropriately identify skin conditions because of their seemingly identical clinical presentations. Thus, this work suggests an embedded vision system based on a CNN model that is put into practice on a low-cost, low-power platform and intended for the real-time monitoring and identification of autoimmune skin conditions. A custom dataset for scleroderma and healthy skin images was created. To enhance diversity, images of vitiligo, psoriasis, and lupus from DermNet website were added due to the small size of the original dataset. To increase accuracy, transfer learning and fine-tuning methods were applied and six state-of-the-art networks were tested: EfficientNet, Xception, Mo-bileNetV2, VGG19, InceptionV3, and ResNet50. All architectures obtained promising results, also thanks to transfer learning, with MobileNetV2leading the way (with a testing accuracy of 95.18% and 75.90% for binary and multi-class classification respectively, and a memory occupancy of 2.5 MB), making it suitable for implementation on a low-cost embedded device, the OpenMV Cam H7 Plus, that may be used to realize an early warning system prototype for autoimmune skin diseases detection.