Enhancing Breast Cancer Image Classification: Minimizing Features with Optimized Feature Extraction and Transfer Learning Techniques
Jay Kshirsagar, B. D. Phulpagar, Pramod Patil · 2024
Computer-aided classification of mammography images plays a crucial role in early and accurate diagnosis of breast cancer. In recent years, transfer learning (TL) techniques have significantly advanced the field of classification, particularly in mitigating the curse of dimensionality. This research work focuses on developing an algorithm for extracting a minimal feature set from breast cancer image datasets, utilizing state-of-the-art technique such as Particle Swarm Optimization (PSO). To address the challenges of feature selection and classification accuracy, our proposed approach employs the ResNet18 and VGG16 architectures for initial feature extraction, generating deep features with enhanced representational capabilities. Subsequently, meta-heuristic algorithms including PSO is employed to further refine and optimize the extracted features, aiming for improved discriminatory power in breast cancer mammography images. The comparative analysis involves assessing TL algorithms to gauge how optimized deep features affect classification performance. The study employs a widely accessible benchmark dataset for experimentation. Significantly, the findings underscore the efficacy of the suggested method, notably showcasing an outstanding F-score of 92.7% by VGG16 and 98.15% with the utilization of ResNet18 with PSO Feature Optimizer. This investigation adds to the continuous endeavors aimed at refining the precision and efficacy of breast cancer diagnosis.