Automated aneurysm detection using machine learning improved by Class-Activation Maps
Mohammad Bhurwani, Kenneth Snyder, Alexander Podgorsak, Ciprian Ionita, Adnan Siddiqui, Ryan Rava, Jason Davies, Elad Levy
Introduction: Convolutional neural networks (CNN) are a type of machine-learning algorithm that can automate the quantitative assessment of intracranial aneurysms (IAs). Class-activation maps (CAMs) can be used to visualize which image regions trigger a trained CNN for different predictions, thus lending insight into how a CNN makes a decision.
Objective: This work investigated the use of a CNN for automatic IA segmentation and radiomic feature extraction for the quantitative assessment of IAs.
Methods: Three hundred and fifty angiographic images of pre- and post-coiled IAs were retrospectively collected. The IAs were manually contoured, and the angiographic sequences and masks were put to a CNN tasked with IA segmentation. CAMs were output to visualize the most salient aneurysmal features. IA segmentation accuracy was assessed with a receiver operating characteristic curve (ROC) using the test cohort. Radiomic features computed with a human user were compared with those computed inside the network IA prediction.
Results: CAMs indicated the IA boundary region is more predictive for segmentation than the interior region. The mean area under the ROC curve for the IA segmentation averaged over the testing cohort was 0.798 (95% confidence interval 0.747-0.824). All five radiomic features measured inside the network IA prediction were within 15% of those measured inside the human contoured IA region.
Conclusions: Automatic segmentation and quantitative assessment of IAs with a CNN has been demonstrated. CAMs can aid in understanding of CNN’s segmentation decisions. The fine-tuning of algorithms and image preprocessing based on these results may improve IA predictive models.