Comparing Different Pre-Trained Models Based on Transfer Learning Technique in Classifying Mammogram Masses

Computer-aided diagnosis (CADx) systems have been applied successfully to many tasks such as diagnosing breast cancer. These systems use classification algorithms to differentiate between malignant and benign lesions based on features extracted from the image data. Recently, more advanced models such as deep convolutional neural networks (CNNs) have been developed that learn features directly from full images; this direct learn process yields fine-grained, hidden features that contain more information than do analytically hand-crafted extracted features. However, building a custom deep-learning model requires extensive computational resources and large amounts of labelled training data. To solve this problem, transfer learning capitalizes on existing rich deep learning models, allowing model creation with significantly reduced training data and time. This paper discusses the potential usefulness of transfer learning for CADx breast cancer task. This is done through investigating many pre-trained models and different fine-tuning techniques to find the optimal model proper for each breast tissue category (Fatty and Dense). We believe that the performance gain from each fine-tuned model specific to breast density is much better for distinguishing benign and malignant masses.