SkinIncept: an ensemble transfer learning-based approach for multiclass skin disease classification using InceptionV3 and InceptionResNetV2

Skin diseases are a significant public health challenge in Bangladesh, with prevalence rates soaring from 11.16 to 63% in recent years. The lack of access to dermatological expertise and resource constraints in rural areas exacerbate delayed or inaccurate diagnoses, leading to worsening conditions and higher treatment costs. This study addresses this critical issue by developing a robust and accurate system for classifying Bangladesh’s ten most common skin diseases using convolutional neural networks (CNNs)-based transfer learning models. Six pre-trained CNN models are implemented, and a novel ensemble model (i.e., SkinIncept) is proposed. Data collection incorporates primary data from the Damien Foundation Hospital and the Bangladesh Institute of Dermatology, STD, and AIDS (BIDSA), supplemented by additional images from reliable web portals. Before model training, extensive preprocessing techniques such as cropping, resizing, filtering, contrast enhancement, histogram equalization, CLAHE, gamma correction, segmentation, and data augmentation are applied to ensure the quality and consistency of the images. The quality of processed images is validated using statistical methods, including MSE, PSNR, SSIM, and RMSE. The performance of each model is rigorously evaluated using several performance metrics, instilling confidence in the study’s methodology and results. Among the six pre-trained models, InceptionResNet-V2 achieved 93.65% accuracy, and the proposed ensemble model achieved a remarkable classification accuracy of 96.52% based on six ablation studies to fine-tune the model and optimize the hyperparameters. These findings form the foundation for “Skin Medicare,” a mobile application providing AI-driven, accessible, and accurate skin disease diagnosis, offering a scalable solution to improve healthcare delivery in underserved and resource-constrained regions.