Enhanced Multi-Class Skin Cancer Classification Using Attention-Based Convolutional Neural Network with HOG Feature Extraction and ROI-Centred Pre-processing

Skin cancer, a common increasingly prevalent disease, encompasses eight distinct diagnostic categories, posing significant challenges for dermatologists due to the highly similar visual characteristics of these categories. Traditional diagnostic accuracy for multiclass skin cancer detection ranges between 70% and 99%, often falling short in multi-class scenarios compared to binary classification performance. To address this, an Attention-Based Convolutional Neural Network (CNN) designed this research for multiclass skin cancer classification. The proposed work begins with a robust pre-processing stage involving Min-Max scaling for normalization and Region of Interest (ROI) cropping, which effectively reduces computational complexity. RGB images are transformed into single-channel grayscale, and ROI cropping emphasizes relevant regions, aiding in the model’s localization capability. Following feature extraction, Histogram of Oriented Gradients (HOG) is used to capture important material details, improving the quality of input features. The classification stage utilizes the Attention based CNN to focus on diagnostically relevant areas, thereby improving classification accuracy across multiple skin cancer types. The proposed method demonstrates a significant improvement in diagnostic performance, offering a promising solution for efficient and accurate cancer classification. The accuracy of the proposed method is achieved as 96% and the specificity is of 0.92.