A comparison of deep learning algorithms for medical image classification and image enhancement

Abstract

In recent years, machine learning techniques based on neural networks have gained popularity. This is primarily because of improved computational capabilities and the availability of larger datasets. In our work, we investigate the application of machine learning techniques, specifically Convolutional Neural Networks (CNNs), for the purpose of medical image analysis. We consider three different tasks for our analysis. Two of these are classification tasks and the third task is an image enhancement task. In the first task, we classify thyroid nodules as malignant versus benign on B-mode and Shear Wave Elastography (SWE) images. We obtain accuracies ranging from 80% - 87% using our evaluated approaches. In the second task, we automatically classify breast Magnetic Resonance Imaging (MRI) images into lesions present and lesion absent classes. For this task, we obtain accuracies ranging from 56% - 69%. In the third project, we train and evaluate a deep learning algorithm for up sampling low resolution ultrasound images and present promising results for obtaining high resolution images from lower quality acquisitions. In general, this work demonstrates that models reliant on deep learning with 104 to 108 unknown parameters can be trained and effectively applied with modest data set sizes on the order of 500 to 10,000 images.