Deep Learning Approach for Optimal Segmentation and Classification of Multi-Class Skin Cancer

This study proposed an efficient image fusion method for skin cancer diagnosis which combines two methods: image segmentation and image enhancement. A new image is created by fusing the k-means clustering and gray scale pseudo-color transformation methods for images. To extract both fine and coarse information from an image, the basic DWT is first used. With DWT, the pictures are broken down into wavelet co-efficients. The approximation coefficient, horizontal detailed coefficient, vertical detailed coefficients, and diagonal detailed coefficients are the four coefficient matrices that are created following decomposition. A fused coefficient matrix is produced by fusing the wavelet coefficients of two images together. The finished image is then recreated using the Inverse Wavelet Transform. The last step is to use the Random Forest Classifier to classify the test features. A number of performance metrics have been employed for evaluating the classification of benign and malignant skin cancer types: accuracy, mean absolute error (MAE), root-mean squared error (RMS), relative absolute error (RAE), root relative squared error (RRS), TP rate, FP rate, precision, recall, F-Measure, MCC, ROC area, and PRC area. With a kappa value of 0.73, MAE of 0.17, RMS value of 0.27, RAE value of 55.60, and RRS value of 69.08, the suggested methodology achieved an accuracy of 92.5%, according to the data. Additionally, the results for the benign class are 0.98, 0.325, 0.92, 0.98, 0.95, 0.75, 0.92, and 0.97 for TP, FP, Precision, Recall, F, MCC, ROC, and PRC, respectively. In contrast, the malignant class yielded results for TP, FP, Precision, Recall, F, MCC, ROC, and PRC of 0.67, 013, 0.893, 0.67, 0.78, 0.75, 0.92, and 0.83, respectively.