Automated, anatomy-based, heuristic post-processing reduces false positives and improves interpretability of deep learning intracranial aneurysm detection models

Deep learning (DL) models can help detect intracranial aneurysms on CTA, but high false positive (FP) rates remain a barrier to clinical translation. We developed a fully automated method to reduce FP by integrating automated in-scan anatomic segmentation and removal of background and venous voxels into hybrid heuristic-DL pipelines. Two DL models, CPM-Net, and a deformable 3D convolutional neural network-transformer hybrid (3D-CNN-TR), trained with 1,186 open-source CTAs (1,373 annotated aneurysms) were integrated into an automated pipeline with heuristic post-processing modules comprising 5 combinations of an open-source brain mask and novel DL-based artery-vein separation modules which create artery, vein, and cavernous venous sinus (CVS) segmentation masks from unlabeled CTA data. FPs that do not overlap with the brain mask and/or overlap with vein or vein-more-than-artery masks were eliminated. Each pipeline was tested on 143 held-out private and 843 publicly available CTAs with 218 and 1027 annotated aneurysms, respectively. On our private dataset, CPM-Net yielded 139 true-positives (TP), 79 false-negative (FN), 126 FP, while 3D-CNN-TR yielded 179 TP, 39 FN, 182 FP. FPs were commonly extracranial (CPM-Net 27.3%; 3D-CNN-TR 42.3%), venous (CPM-Net 56.3%; 3D-CNN-TR 29.1%), arterial (CPM-Net 11.9%; 3D-CNN-TR 53.3%), and non-vascular (CPM-Net 25.4%; 3D-CNN-TR 9.3%) structures. Method 5 (combination of brain and vein-more-than-artery mask) performed best, reducing FP by 70.6% (89/126) and 51.6% (94/182) without reducing TP, lowering the FPR from 0.88 to 0.26, and from 1.27 to 0.62 for CPM-Net and 3D-CNN-TR, respectively. On the public RSNA dataset, CPM-Net yielded 791 TP, 236 FN, 748 FP, 0.89 FPR; while 3D-CNN-TR yielded 940 TP, 87 FN, 1552 FP, 1.84 FPR. Method 1 (enhanced brain mask) performed best at preserving TP, removing 1 and 4 TP for CPM-Net and 3D-CNN-TR, respectively, while removing 31.4% (235/748) and 33.8% (524/1552) of FP. Method 5 eliminated 57.9% and 43.8% of FP, but removed 24 and 25 TP from CPM-Net and 3D-CNN-TR output, respectively. Integration of interpretable, anatomy-based, background and vein removal modules into a fully automated DL-based aneurysm detection pipeline improved model performance on two external test datasets. This suggests that heuristic-DL hybrid pipelines created by integrating in-pipeline domain-informed heuristic post-processing with DL may increase performance and clinical acceptance of radiology domain AI.