R3MV: a novel reliable system architecture for skin cancer classification using progressive heterogeneous multiblock model

Abstract Medical picture categorization has been greatly enhanced by the use of deep learning, especially in the timely identification of skin lesions. Still, predictions from a single model remain unreliable due to their susceptibility to variations in the dataset, complicating their application to diverse clinical scenarios. This study introduces a unique CNN, PHMBCNN, designed to enhance classification accuracy using a progressive learning strategy. We propose the R3MV three-tier decision fusion system, which integrates predictions from (i) individual CNN model predictions, (ii) a feature fusion classification architecture, and (iii) a meta-classifier trained on the outputs of the CNN models. The final forecast is reached using a majority voting procedure, which enhances the reliability of the decision. The study utilizes two datasets for skin cancer: PAD_UFES_20 and HAM10000. Incorporating a GRU into the PHMBCNN model results in the PHMBCNN-GRU. The classification accuracy enhanced from 75.70% for the PAD_UFES_20 dataset to 80.69%, and from 92.07% for the HAM10000 dataset to 96.01%. The R3MV system design achieves 81.78% for PAD_UFES_20 and 99.3% for HAM10000.