Optimization of DSA image data input to a machine learning aneurysm identifier

Convolutional neural networks (CNN) can automate the quantitative assessment of intracranial aneurysms (IAs); however, as a “black-box” technique, it does not allow users to understand which image features are most important, and thus how to improve network predictions. Class-activation maps (CAMs) are used to visualize which image regions trigger a trained CNN, thus lending insight into how a CNN makes a decision. This work investigated the use of CAMs to identify differences in network activation inside IAs for an image segmentation task for the goal of optimizing the pre-processing framework. Three hundred and fifty angiographic images of pre- and post-coiled IAs were retrospectively collected. The IAs were manually contoured, angiographic sequences were flattened along the time-axis using different techniques (average, median, standard deviation, or minimum value), and flattened sequences and masks were put to a CNN tasked with IA segmentation. CAMs were output to visualize the most salient aneurysmal features. Network activation was higher in the IA peripheral region compared with the IA middle region indicating the IA periphery is more predictive for segmentation. Flattening angiographic sequences using the average value along the time-axis leads to the most accurate IA segmentations with an average Dice coefficient of 0.765 and an average Jaccard Index of 0.624 over the test cohort. This work indicates that CAMs can aid in understanding of a CNN's segmentation decisions. Fine-tuning and automation of algorithms along with network-input image preprocessing based on these results may improve IA predictive models.