Deep feature extraction and optimized VGG16-SVM architecture for breast cancer characterization

This research endeavours to devise a novel model that can achieve optimal accuracy levels, even in scenarios with insufficient images available for extensive training and testing, as demonstrated by the Breast Ultrasound Image (BUSI) dataset. In the first stage, the pre-processed images, which were subjected to resizing and re-scaling, were introduced to the proposed models. Then, various train, validation, and test split ratios were explored to obtain the best-performing model. The initial experiment emphasized gaining insight into the selected dataset; hence, the CNN model was applied and further optimized with RGB and scaled images, using an 81/9/10 train, validation, and test split ratio, which outperformed others with 76.92% accuracy. The 9knowledge derived from this best-performing model guided subsequent experiments. Consequently, by employing a Sequence Learning (SL) approach on CNN, featuring Transfer Learning (TL) with diverse Support Vector Machine (SVM) classifiers, the refined model achieved 80.12% accuracy. Notably, during hyperparameter tuning, the Regularization (C-value) and K-fold cross-validation (cv value) were established at 500 and 25, respectively, considering insights from the previous experiment. Subsequently, applying SL to the more complex VGG16 model with TL and optimized SVM, the best-performing model achieved an impressive accuracy of 91.02%. The result gained from prior experiments plays a significant role in the success of the VGG16-SVM model, specifically in hyperparameter fine-tuning, with C = 100. This optimized model surpassed all other experiments in this study and exceeded the results of earlier studies for the BUSI dataset.