Beyond Fine-tuning: Classifying High Resolution Mammograms using Function-Preserving Transformations

The task of classifying mammograms is very challenging because the lesion isusually small in the high resolution image. The current state-of-the-artapproaches for medical image classification rely on using the de-facto methodfor ConvNets - fine-tuning. However, there are fundamental differences betweennatural images and medical images, which based on existing evidence from theliterature, limits the overall performance gain when designed with algorithmicapproaches. In this paper, we propose to go beyond fine-tuning by introducing anovel framework called MorphHR, in which we highlight a new transfer learningscheme. The idea behind the proposed framework is to integratefunction-preserving transformations, for any continuous non-linear activationneurons, to internally regularise the network for improving mammogramsclassification. The proposed solution offers two major advantages over theexisting techniques. Firstly and unlike fine-tuning, the proposed approachallows for modifying not only the last few layers but also several of the firstones on a deep ConvNet. By doing this, we can design the network front to besuitable for learning domain specific features. Secondly, the proposed schemeis scalable to hardware. Therefore, one can fit high resolution images onstandard GPU memory. We show that by using high resolution images, one preventslosing relevant information. We demonstrate, through numerical and visualexperiments, that the proposed approach yields to a significant improvement inthe classification performance over state-of-the-art techniques, and is indeedon a par with radiology experts. Moreover and for generalisation purposes, weshow the effectiveness of the proposed learning scheme on another largedataset, the ChestX-ray14, surpassing current state-of-the-art techniques.