Deep learning for medical image analysis

Abstract

Deep learning has the capability to learn features in images and classify them in supervised tasks. There are many parameters to a deep learning model, providing immense flexibility and a high degree of customisability to each task. However, such freedom can be a double-edged sword as it becomes a challenge to find the best set of values for each parameter among a myriad of others. In the context of medical images, finding a set of methods to boost performance can enable deep learning methods to effectively assist medical professionals in making diagnoses. This report details the methods taken to explore the effects of spatial attention mapping, loss function variation, the use of pretrained layers and layer freezing, and batch size variation on deep learning performance on medical images. Experiments were conducted for each parameter on two medical image datasets. It was concluded that Spatial Attention Mapping can significantly boost the accuracy performance of deep learning while the use of BCE Logit Loss function and Focal Loss function is dependent on the nature of the images in the dataset and possibly the proportion of classes. Pretrained layers and layer freezing are very effective when used in combination, and batch size variation provides very little discernible performance difference but can be tuned to reduce training times.