Enhancing Melanoma Skin Cancer Detection Through Feature Fusion of Pre-Trained Deep Convolutional Neural Network ResNet50 and Thepade Sorted Block Truncation Coding

• Published in: SN computer science 2024-04, Vol.5 (4), p.426, Article 426
• Main Authors: Thepade, Sudeep D., Shukla, Sudhanshu
• Format: Article
• Language: English
• Subjects: Accuracy; Algorithms; Artificial neural networks; Cancer; Classifiers; Clinical outcomes; Coding; Computer Imaging; Computer Science; Computer Systems Organization and Communication Networks; Data Structures and Information Theory; Datasets; Deep learning; Effectiveness; Feature extraction; Genetics; Information Systems and Communication Service; Literature reviews; Machine learning; Medical diagnosis; Medical prognosis; Melanoma; Methods; Mutation; Neural networks; Original Research; Pattern Recognition and Graphics; Skin cancer; Software Engineering/Programming and Operating Systems; Squamous cell carcinoma; Vision
• ISSN: 2661-8907; 2662-995X; 2661-8907
• DOI: 10.1007/s42979-024-02821-5

Identifying melanoma skin cancer accurately and promptly is crucial for successfully applying treatment methods. While helpful, conventional methods often must catch up to the desired accuracy level. This paper presents a pioneering approach to overcome these limitations. The crux of our investigation was assessing the effectiveness of pre-trained deep convolutional neural network (DCNN) models, especially ResNet50, when merged with Thepade sorted block truncation coding (Thepade SBTC) for detecting Melanoma. We aimed to develop and evaluate a feature fusion technique incorporating these methodologies to improve diagnostic accuracy. The research also aimed to compare this newly suggested approach with existing methods, appraise its performance via appropriate metrics, and elucidate the potential advantages and limitations. To realise these objectives, the work proposed here combined the pre-trained DCNN model ResNet50 and Thepade SBTC features. These models were trained and evaluated using the HAM10000 datasets. Subsequently, we combined the features extracted from these models with Thepade SBTC 10-ary features to enhance the machine learning classifiers' discernment abilities. We further fine-tuned the pre-trained models to adapt them explicitly for cancer detection. Interestingly, integrating CNN features and Thepade SBTC 8-ary improved the Random Forest classifier's effectiveness, achieving an accuracy rate of 92.78%. A suite of RandomForest, IBK, and NBTree classifiers combined with CNN features and Thepade SBTC 10-ary yielded an accuracy of 92.35%. The findings indicate that merging ResNet50's deep learning features with Thepade SBTC-ary attributes significantly improves melanoma skin cancer detection, albeit slightly increasing the computational demands for feature extraction. Thus, this study contributes to developing more adept and dependable methods for the early detection and treatment of melanoma skin cancer, signposting a promising direction for further research.