Evaluation of the impact of deep learning architectural components selection and dataset size on a medical imaging task

Deep Learning (DL) has been successfully applied in numerous fields fueled by increasing computational power and access to data. However, for medical imaging tasks, limited training set size is a common challenge when applying DL. This paper explores the applicability of DL to the task of classifying a single axial slice from a CT exam into one of six anatomy regions. A total of similar to 29000 images selected from 223 CT exams were manually labeled for ground truth. An additional 54 exams were labeled and used as an independent test set. The network architecture developed for this application is composed of 6 convolutional layers and 2 fully connected layers with RELU non-linear activations between each layer. Max-pooling was used after every second convolutional layer, and a softmax layer was used at the end. Given this base architecture, the effect of inclusion of network architecture components such as Dropout and Batch Normalization on network performance and training is explored. The network performance as a function of training and validation set size is characterized by training each network architecture variation using 5,10,20,40,50 and 100% of the available training data. The performance comparison of the various network architectures was done for anatomy classification as well as two computer vision datasets. The anatomy classifier accuracy varied from 74.1% to 92.3% in this study depending on the training size and network layout used. Dropout layers improved the model accuracy for all training sizes.