Influence of input image configurations on output of a convolutional neural network to detect cerebral aneurysms

Rupture of cerebral aneurysms is the main cause of subarachnoid hemorrhage, which can have devastating effects on quality of life. The identification and assessment of unruptured aneurysms from medical images is therefore of significant clinical relevance. In recent years, the availability of clinical imaging data has rapidly increased, which calls for computer assisted detection (CAD) systems. Previous studies have shown that CAD systems based on convolutional neural networks (CNN) can help to detect cerebral aneurysms from magnetic resonance angiographies (MRAs). However, these CAD systems require large datasets of annotated medical images. Thus, more efficient tools for processing and categorizing medical imaging data are required. Previous studies of CNN-based classification for medical images used various patch configurations of input data. These studies showed that classification accuracy was affected by the patch size or image representation. Thus, we hypothesize that the accuracy of CADs to detect cerebral aneurysms can be improved by adjusting the configuration of the input patches. In the present study, we performed CNN-based medical imaging classification for varying input data configurations to examine the relationship between classification accuracy and data configuration.