A Novel Deep Learning Based Melanoma Detection Technique Using Skin Lesion

A 'Melanoma'is a type of skin cancer with the highest fatal rate. It accounts for up to 1.7% of cancer diagnoses globally. Early detection of the disease has many advantages, it significantly reduces the chances of morbidity. The therapies involved are very critical and has an adverse effect on patients mental and physical condition, usually biopsy is advised for detection. The procedure itself is very hard and gives difficult stretch to a patient. In addition, the patient needs to wait for around 3–12 weeks to get results. The procedures used in the earliest form of melanoma detection are believed to be expensive and somewhat inaccurate. In this article, a novel deep learning-based approach is proposed to diagnose the disease in a timely manner. Currently due to less fragmented data the existing solutions has not been adopted by a practical industry. During the exploratory data analysis, it has been found that data is highly imbalance. To mitigate this lack of proportion the calculated weightage for each class has been utilised. In order to oversee the impact of this weightage we have reduced the count of classes which have higher data to not more than 4-folds when comparing to classes containing lower data count. In the combination we have used EfficientNet-B1 for classification along with more hidden and fully connected layers. The proposed models have been evaluated through industry practices. The 86% accuracy was acheived with the process. Cross validation on the model approach has been applied to find unbiased result set of accuracy, precision, recall and F1-score. To be more discrete, 10-folds are used on the proposed deep learning model approach. Previously an accuracy of 81% (Res-Net) and 82% (Dense-Net) were reported for the same models. The dermoscopic images are highly complacent, after shedding some more light it was evident that the inner attributes of the image play a vital role and provide valuable insights to the model. The rotation range, zoom range and rescale is being focused as they sharpen the object edges and help to identify the sheer pointers prominently. The complete data is provided to the models which significantly improves the results.