Development of Dermatological Lesion Detection System Using EfficientNet with Fairness Evaluation

The most common diseases in the world are skin problems because of hereditary features and environmental factors. It can be very challenging to distinguish between different skin illnesses and their different types. Given the complexity of human skin complexion and the visual proximity effect on illnesses, accurately identifying the exact type of diseases can be challenging. The intricacies of skin conditions, coupled with variations in skin tones, make it difficult to pinpoint specific diseases with certainty. Therefore, ii is crucial to recognize and classify skin problems as soon as they are found. In the biomedical sector, ML and DL techniques are extensively used for segmentation, augmentation, feature extraction, classification, diagnosis, and detection in its early stage. This work presents an automated deep learning-based image-based transfer learning system for recognizing and categorizing skin cancer. ESRGAN is used for image enhancement. Furthermore, for the feature extraction and classification of the skin lesion into seven classes from the HAM10000 dataset, a pre-trained EfficientNet model has been used. The classification results of the suggested model have been shown in accuracy, recall, AUC, F1-score, precision, and ROC. Also, the performance is evaluated using fairness evaluation metrics. This procedure starts with a digital photo of the sick skin region and then analyzed to determine the kind of illness discovered. Since this research provides results more quickly and accurately than the previous procedures, it will be a more reliable and effective strategy than the conventional method for identifying dermatological problems.