Classification of Skin Lesions to Predict Melanoma Using Deep Learning Algorithm on ESP32

This work describes a computer image-based deep learning system for identifying and classifying skin cancer that is developed using an ESP-32 microcontroller. We train and test our CNN model using a dataset of over 10,000 images obtained from HAM10000. This research classifies skin cancer into seven kinds. This research addresses the most prominent skin cancer types as classes such as melanoma, basal cell carcinoma, benign keratosis, dermatofibroma, nevus, actinic keratosis, and vascular lesions. CNN algorithm is used to perform the detection which is known for its feature extraction and ease the use of large data which in this case is various types of skin cancer which is processed in the training model. In addition to CNN, we suggested a novel approach that uses ViT-B16, ViT-B32, EfficientNet V2, ResNet-50 V2, and ensemble methods to enhance prediction results on skin cancer diagnosis. We suggest using ViT-B16 and ViT-B32 in combination with EfficientNet V2 and ResNet-50 V2, a cutting-edge deep learning architecture, that extracts features from skin lesion images. To increase the classification model's performance, three distinct ensemble approaches are used: weighted average ensemble, geometric mean ensemble, and average ensemble. The collected characteristics are then loaded into a binary classification model, which uses all of the approaches discussed above to discriminate between melanoma and benign skin lesions. Our CNN model surpassed prior recent techniques, achieving an accuracy of 83 % percent when deployed on ESP-32. Our findings show that deep learning-based approaches applied to microcontrollers can reduce computing costs and develop a portable/trained device that predicts skin cancer based on test data.