Studies applying artificial intelligence as a reader-aid for the detection of cerebral aneurysms
| Author | Modality | Demographics: total cases; aneurysm-positive cases (+); total aneurysms (A). If data available: % female; mean age; mean aneurysm diameter (range); rupture status | Study design | Dataset: number of cases (number of which contain aneurysms (+), if different from the total number of cases) | Reference | Index test used | Type of test set | AI performance: if available: lesion sensitivity; FP/case; accuracy; patient sensitivity; patient specificity; AUC | Reader performance: If available: lesion sensitivity; FP/case; accuracy; patient sensitivity; patient specificity; AUC | Reader+AI performance: If available: lesion sensitivity; FP/case; accuracy; patient sensitivity; patient specificity; AUC | Position of AI in pipeline |
| Miki et al 201655 | MRA-TOF | 2701 cases; 189+; 203A; 38% female; 54 years (median) | Prospective single center | Test=2701 (189+) | 2 radiologists reviewed MRA, with CAD assistance | cCAD: Gaussian filter based on Nomura et al 201424 (see table 1) | Internal; temporal split | Lesion sensitivity: 82% | Lesion sensitivity: 64% | Lesion sensitivity: 69% | 2nd reader |
| Štepán-Buksakowska et al 201456 | MRA-TOF | 48 cases; 9+; 11A; 3.1 mm | Retrospective multicenter | Test=48 (9+) | 2 neuroradiologists reviewed MRA-TOF and DSA | cCAD: Dot enhancement filter based on Yang et al 201126 (see table 1) | External; Geographical split | Lesion sensitivity: 91% | Patient sensitivity: 70.4%; specificity: 79.5%; accuracy: 77.8%; AUC: 0.66 | Patient sensitivity: 83.4%; patient specificity: 75.7%; accuracy: 77.1%; AUC: 0.76 | 1st reader |
| Miki et al 202057 | MRA-TOF | 250 cases; 100+; 104A; unruptured | Retrospective single center | Test=250 (100+) | 2 radiologists reviewed MRA-TOF | DL: Voxel based CNN based on Nakao et al 201828 (see table 1) | Internal; hold out | Lesion sensitivity: 92.3% | NR | Lesion sensitivity: 71%; lesion specificity: 92%; AUC 0.89 | Unclear |
| Kakeda et al 200858 | MRA-TOF | 50 cases; 16+; 16A; 54% female, 64 years; 6.6 mm (3–26 mm); unruptured | Retrospective multicenter | Test=50 (16+) | 2 neuroradiologists reviewed MRA and other available imaging | cCAD: 3D selective enhancement filter using Hessian matrix based on Arimura et al 200623 (see table 1) | Internal; LOOCV | Lesion sensitivity: 81% FP/case: 2.7 | Patient sensitivity: 75.8%; patient specificity: 82.7%; AUC: 0.85 | Patient sensitivity: 80.9%; patient specificity: 88.6%; AUC: 0.90 | 2nd reader |
| Hirai et al 200559 | MRA-TOF | 50 cases; 22+; 68% female; 59 years; 7.1 mm (3–26 mm); unruptured | Retrospective single center | Test=50 (22+) | 2 neuroradiologists reviewed MRA and other available imaging | cCAD: 3D selective enhancement filter using Hessian matrix based on Arimura et al 200424 (see table 1) | Internal; temporal split | NR | AUC: 0.93 | AUC: 0.98 | 2nd reader |
| Sohn et al 202160 | MRA-TOF | 332 cases; 135+; 169A; 64% female, 62 years; 4 mm (2–17 mm); unruptured | Retrospective single center | Test=332 (135+) | 3 neuroradiologists reviewed MRA other available imaging | DL: 3D ResNet based on Joo et al 202025 (see table 1) | Internal; temporal split | Lesion sensitivity: 92.3%. FP/case: 0.12; patient sensitivity: 74.8%; specificity: 93.9% | Lesion sensitivity: 74.8%; patient sensitivity: 73.5%; patient specificity: 94.8% | Lesion sensitivity 95.0%; patient sensitivity: 86.5%; patient specificity 95.2% | 2nd reader |
| Park et al 201961 | CTA | 818 cases; 328+; 358A; 64% female; 58 years; unruptured | Retrospective single center | Training=611 (223+); Validation=92 (46+); Test=115 cases (56+) | 1 neuroradiologist reviewed CTA and DSA (if available) | DL: 3D CNN HeadXNet | Internal; hold out | Patient sensitivity: 94.9%; patient specificity: 66.1%; accuracy: 80.9% | Patient sensitivity: 83.1%; specificity: 96.0%; accuracy: 89.3% | Patient sensitivity: 89.0%; patient specificity: 97.5%; accuracy: 93.2% | 1st reader |
| Pennig et al 202162 | CTA | 172 cases; 172+; 205A; 63% female; 55.4 years; ruptured and unruptured | Retrospective single center | Training and validation=68; Test=104 | 1 neurosurgeon and 1 radiologist reviewed CTA and DSA (if available) | DL: ensemble model based on Shahzad et al 202051 (see table 1) | Internal; temporally distinct | Lesion sensitivity: 85.7%; FP/case: 0.84 | Lesion sensitivity: 88.1% | Lesion sensitivity: 97.1% | 2nd reader |
| Yang et al 202163 | CTA | 1468 cases; 1256+; 1543A; 48% female; 57 years; 4.1 mm (1–22 mm); ruptured and unruptured | Retrospective multicenter | Training=534; Validation=534; Test=400 (188+) | 2 radiologists reviewed CTA | DL: 3D CNN ResNet-18 | Internal; temporally distinct | Lesion sensitivity: 97.5% FP/case: 13.8 (internal validation set) | Lesion sensitivity: 79.1%; patient sensitivity: 81.6%; specificity: 95.9%; AUC: 0.60 | Lesion sensitivity: 88.9%; patient sensitivity: 91.9%; specificity: 90.9%; AUC: 0.61 | Unclear |
+, aneurysm-positive cases; A, aneurysm; AI, artificial intelligence; AUC, area-under-curve; CTA, CT angiography; CV, cross-validation; 2D, two dimensional; 3D, three dimensional; DL, deep learning; DSA, digital subtraction angiography; FP/case, false-positives per case; LOO, leave-one-out; MRA-TOF, MR angiography-time of flight; NR, not recorded.