Portia Amoako Tawiah

Basildon and Thurrock University Hospital, United Kingdom

Abstract Title:

Cancer or not: Artificial intelligence for skin cancer detection

Research Interests:

Introduction: Skin cancer is one of the most common malignancies globally, with melanoma cases alone affecting over 15,000 individuals annually in the UK. Early detection is critical however increasing incidence rates place significant pressure on healthcare systems. Artificial intelligence, particularly Convolutional Neural Networks (CNNs), has emerged as a promising solution for enhancing diagnostic accuracy and efficiency.

Objective: This study aimed to develop an AI-based skin cancer detection model using the ResNet50 architecture and transfer learning. Methods The HAM10000 dataset, supplemented with SCC data from the ISIC archive, was pre-processed, standardised, and split into training and validation sets. Data augmentation and class weighting addressed class imbalance, and transfer learning enabled fine-tuning of the ResNet50 model.

Results:  Model achieved an overall accuracy of 79.76% in training and 77% in validation, with robust performance for Squamous Cell Carcinoma (SCC), achieving a precision of 0.89 and an F1-score of 0.84. The confusion matrix showed excellent performance for dominant classes like Nevus (Class 0) with 1,089 correct predictions, but reduced performance for minority classes such as Benign Keratosis-Like Lesions (Class 2) and Dermatofibroma (Class 6), highlighting class imbalance. ROC curves indicated excellent discriminatory ability, with AUC values ranging from 0.90 to 1.00. Notably, Squamous Cell Carcinoma (Class 7) achieved high precision (0.89) and F1-score (0.84), reflecting effective class-weighting techniques.

Conclusion: While the model performed well for most classes, underrepresented categories, such as Benign Keratosis-Like Lesions and Dermatofibroma, exhibited reduced performance due to class imbalance. Ensuring dataset diversity and using advanced augmentation techniques could enhance performance. These findings demonstrate the potential of CNNs for skin cancer detection and highlight the need for further refinement to optimise clinical integration.