Skin Multi Class Cancer Classification Using ResXceptNet with Layer Wise Integration

Skin cancer is considered as one of the most prevalent and deadly types of cancer which still remains a complex task for dermatologist during classification.The complexity arouses due to the smaller regions and boundaries of the cancer affected partitions.The existing techniques give rise to poor classification accuracy due to the poor performance of the classifier in recognizing complex patterns.So, an effective classification method is proposed with the combinational classifier of ResNet 50V2 and Xception Net namely, ResXceptNet.The ResXceptNet capacity is enhanced by obtaining complex and subtle patterns which additionally help overcome the overfitting risk, as well as designing the robust and generalizable classification.Next, the GradCAM guided attention, guided backpropagation and Layer-wise Relevance Propagation (LRP) are incorporated in ResXceptNet to enhance the ability of the model to classify complex skin lesion patterns.The pre-processing and data augmentation are performed over HAM10000, ISIC-2019 and ISIC-2020 datasets for enhancing and generating the diverse variation in data.Here, the VGG-19 is utilized in extracting appropriate features to perform multi class classification using ResXceptNet.The performance of ResXceptNet is analysed using accuracy, precision, recall and F-measure.The existing approaches such as Convolutional Neural Network-Discrete Wavelet Transformation (CNN-DWT), CNN-Local Binary Pattern (CNN-LBP) and Skin-Net are considered for evaluation of ResXceptNet.The accuracy of ResXceptNet is 99.17%, 98.54% and 99.45% respectively for HAM10000, ISIC-2019 and ISIC-2020 datasets, proving superior to CNN-DWT and CNN-LBP.