MULTICLASS SKIN CANCER CLASSIFICATION USING RESNET AND DERMOSCOPIC IMAGING

Abstract

Skin cancer is among the most common and life-threatening illnesses globally, and early and precise diagnosis is necessary to enhance the outcome of treatment. Conventional diagnosis is based largely on skilled dermatologists, which is time-consuming and variable. Over the past few years, deep learning methods have proven to have vast potential in computerizing medical image analysis, both speeding up and improving the accuracy of diagnosis. This work aims to develop a reliable deep learning-based classification model using the ResNet architecture for identifying and categorizing various types of skin cancer. The research makes use of the HAM10000 dataset, an extensive collection of dermoscopic images representing various skin cancer types, to ensure representative and varied data for training and testing. The methodology suggested here encompasses several stages such as dataset preprocessing, feature extraction based on pre-trained deep learning models, and classification based on ResNet. Several data augmentation methods and preprocessing operations are utilized to improve model accuracy and handle imbalances in classes. Feature extraction takes advantage of ResNet's hierarchical feature representation, where complex patterns in dermoscopic images are captured to support precise classification. The model is optimized and fine-tuned with the best hyperparameters to achieve maximum classification accuracy with minimal computational complexity. A comparative analysis with traditional machine learning methods underscores the superiority of deep learning techniques in dealing with sophisticated visual data. Experimental findings show that the model proposed has excellent classification accuracy, performing better than traditional methods in precision, recall, F1-score, and overall classification accuracy. The model also separates various types of skin cancers correctly with minimal false positives and negatives. Performance monitoring methods like cross-validation and regularization are also embedded to prevent overfitting so that the model is generalizable to unseen samples. The computational cost of the method is examined to ascertain its viability for real-time clinical use, highlighting its prospect of being part of automated diagnosis systems.

Keywords

(Deep Learning, Skin Cancer, Skin Cancer Classification, Skin Images).