Transfer Learning in Deep Convolutional Neural Networks for Detection of Architectural Distortion in Digital Mammography

Deep learning models have reached superior results in various fields of application, but in many cases at a high cost of processing or large amount of data available. In most of them, specially in the medical field, the scarcity of training data limits the performance of these models. Among the strategies to overcome the lack of data, there is data augmentation, transfer learning and fine-tuning. In this work we compared different approaches to train deep convolutional neural network (CNN) to automatically detect architectural distortion (AD) in digital mammography. Although several computer vision based algorithms were designed to detect lesions in digital mammography, most of them perform poorly while detecting AD. We used the VGG-16 network pre-trained on ImageNet database with progressive fine-tuning to evaluate its performance on AD detection over a database of 280 images of clinical mammograms. Finally, we compared the results with a custom CNN architecture trained from scratch for the same task. Results indicated that a network with transfer learning and certain level of fine-tuning reaches the best results for the task (AUC = 0.89) compared with the other approaches, but no statistically significant difference was found between the best results using different amount of data augmentation and also compared to the custom CNN.