Modelling of hybrid deep ensemble learning based skin lesion detection using FCNN denoising and inception-dilated ResNetV2.

Specific metabolic and genetic abnormalities can cause skin lesions with cancerous potential. Cancerous cells might be spreading to every part of the body, whereas cancer is critical. Skin cancer is one of the prevalent cancers, and its global incidence continues to rise. Skin lesion classification is a significant stage in computer-aided diagnosis (CAD) for automatic analysis of melanoma. Recently, increased focus has been given to the deep learning (DL) methods applied for image analysis owing to their capability to utilise machine learning (ML) methods to convert input data into higher-level performance. Owing to precise diagnosis, the healthcare domain has a constantly rising interest in this technology, particularly in the analysis of melanoma. In this study, an integration of Residual Learning and Feature Extraction for Skin Lesion Classification (IRLFE-SLC) method is proposed for medical imaging. Initially, a fully convolutional neural network (FCNN) model is used for image denoising to effectively extract noise from dermoscopic images. Next, the feature extraction mechanism is employed using an Inception-dilated ResNetv2 model to capture multi-scale and hierarchical features critical for lesion characterisation. For the skin lesion classification, an ensemble of three DL models, such as bidirectional long short-term memory (Bi-LSTM), deep belief networks (DBN), and spiking neural networks (SNN) models, is utilised. To show the improved performance of the IRLFE-SLC methodology, a comprehensive investigational analysis is conducted. The comparison study of the IRLFE-SLC approach showed a superior accuracy of 99.06% and muti-class AUC score of 95.68% (macro-averaged) compared with existing models under an 80:20 split on the International Skin Imaging Collaboration (ISIC) skin cancer dataset.