Skin cancer is one of the most prevalent types of cancer worldwide, and its early detection is crucial for effective treatment. In this study, we propose a deep learning-based approach to classify skin cancer images using the ResNet50 architecture on the HAM10000 dataset. The dataset consists of images from seven different classes of skin cancer, with the challenge of class imbalance present in the data. To mitigate this, we employed class-weighted loss functions. Our model achieved an impressive validation accuracy of 84.32% after 10 epochs of training. The results demonstrate the effectiveness of ResNet in skin cancer classification, as well as the importance of handling class imbalance. Future work includes experimenting with more advanced architectures and techniques for further performance improvement.
Skin cancer, particularly melanoma, is one of the leading causes of cancer-related deaths globally. Timely diagnosis is essential for increasing survival rates, but traditional diagnostic methods are time-consuming and require expertise. With the advent of deep learning (DL), there is significant potential to automate skin cancer detection and improve diagnostic accuracy. This paper focuses on the development of an automated skin cancer classification system using deep learning, specifically leveraging the ResNet50 model. We use the HAM10000 dataset, a publicly available collection of dermatology images, to train and evaluate the model. One of the challenges in this dataset is the imbalance in the number of images across the seven classes. To address this, we incorporate class-weighted loss functions during model training. The objective of this work is to develop a high-accuracy model capable of classifying skin lesions into one of seven categories, offering a reliable tool for aiding dermatologists in diagnosing skin cancer.
In this study, a ResNet-based deep learning model was developed and evaluated for the classification of skin cancer using the HAM10000 dataset. The system achieved a validation accuracy of 84.32%, demonstrating strong performance in identifying seven distinct types of skin lesions. By incorporating class weighting and data augmentation, the model effectively addressed the issue of class imbalance, resulting in improved prediction consistency across both majority and minority classes.
The proposed approach offers a balanced solution between model complexity and performance, making it suitable for real-world medical applications. Additionally, the inclusion of a user-friendly prediction module allows for practical deployment, enabling real-time classification of dermatoscopic images. Although there are limitations related to data diversity, explainability, and metadata integration, the system serves as a solid foundation for future enhancements.
Overall, this work highlights the potential of deep learning models like ResNet in aiding dermatologists with accurate and automated skin cancer diagnosis. With further development, clinical validation, and integration of additional data sources, the system could contribute meaningfully to early detection and treatment planning, ultimately improving patient outcomes.
In addition to achieving high validation accuracy, the model demonstrated consistent improvements during training, with a significant decrease in training loss from 280.26 in the first epoch to 52.26 by the tenth epoch. The accuracy curve showed steady growth, reaching 87.61% training accuracy, indicating that the model was effectively learning the features of different skin lesion types. The system was trained using two parts of the HAM10000 image dataset and corresponding metadata, with careful preprocessing and structured loading to ensure reliable results.