Breast Cancer Detection by Prototypical Networks using Few Shot Learning

The presented study attempts to use a meta-learning paradigm, based on the Breast Ultrasound Image (BUSI) dataset. With the Protonet convolutional neural network architecture and a novel augmentation methodology, the main objective is to accurately classify breast ultrasonography images into benign, malignant, and normal categories. The utilization of few-shot learning in the meta-learning framework guarantees flexibility and generalization, especially in situations with the scarcity of annotated medical image datasets. Our experimental findings demonstrate the strong effectiveness of the meta-learning strategy by revealing notable improvements in classification performance on the BUSI dataset. Images are processed efficiently using multithreading. This highlights the model's robust generalization to unknown data, securing its potential for significant applications in the identification of breast cancer. Notably, when the number of labeled examples are k = 5, 10, 15 the test accuracies achieved are 85.75%, 91.28%, 91.67%. This model outperforms other CNN models in terms of efficiency.