Automatic classification of breast lesions using Transfer Learning
Abstract
Breast lesions are one of the most common types of lesions among women in Brazil and worldwide, accounting for about 28% of new cases each year. These lesions may have Benign or Malignant behaviors. In this work, a computational method for image classification was developed to differentiate malignant and benign breast lesions, aiming at a low computational cost and good efficiency. In our approach, different Convolutional Neural Networks architectures and several classifiers were tested. Transfer Learning was employed to deal with the limitation of the small number of images in the database, reaching an accuracy of 81.73%, a sensitivity of 85.66%, a specificity of 78.40%, Kappa of 0.63 and ROC curve of 0.82. Finally, we believe our method can integrate a CAD tool by acting as patient screening or by providing a second opinion to the specialist.