Studies applying artificial intelligence as a standalone method for the automatic detection of cerebral aneurysms
| Author | Modality | Study design | Demographics: total cases; aneurysm-positive cases (+); total aneurysms (A). If data available: % female; mean age; mean aneurysm diameter (range); rupture status | Reference standard | Index test | Dataset: number of cases (number of which contain aneurysms (+), if different from the total number of cases) | Type of test set | Test set performance(If available): lesion sensitivity; FP/case; PPV; NPV; accuracy; patient sensitivity; patient specificity; precision; AUC; F1 score |
| Nomura et al 201421 | MRA-TOF | Retrospective single center | 2269 cases; 472+; 578A; 3 mm | 2 radiologists reviewed MRA-TOF | cCAD: Gaussian filter with boosting algorithm (AdaC2) | Training=490 (362+); Test=1779 (110+) | Internal; temporal split | Lesion sensitivity: 95.2%; FP/case: 9 |
| Jin et al 201622 | MRA-TOF | Retrospective single center | 30 cases; 30+; 31A; 69.2 years; 67% female; 3.7 mm (2.0–5.5 mm); unruptured | 2 neuroradiologists reviewed CTA/DSA | cCAD: ellipsoid convex enhancement filter | Test=30 | Internal; LOOCV | Lesion sensitivity: 100%; FP/case: 31.8 |
| Arimura et al 200623 | MRA-TOF | Retrospective multicenter | 178 cases; 87+; 214A; 5.2 mm (1–15 mm); unruptured | 2 neuroradiologists reviewed MRA-TOF and other available imaging | cCAD: 3D selective enhancement filter using Hessian matrix, shape-based difference image technique | Training=115 (53+); Training/test=63 (34+) | Internal; LOOCV | Lesion sensitivity: 94%;, FP/case: 2.3 |
| Arimura et al 200424 | MRA-TOF | Retrospective single center | 60 cases; 29+; 36A; 6.6 mm (3–26 mm); unruptured | NR | cCAD: 3D selective enhancement filter using Hessian matrix | Test=60 (29+) | Internal; LOOCV | Lesion sensitivity: 100%; FP/case: 2.4 |
| Joo et al 202025 | MRA-TOF | Retrospective multicenter | 744 cases; 644+; 761A; 3.2 mm; unruptured | 2 neuroradiologists reviewed MRA-TOF | DL: 3D ResNet | Training and validation=468; Internal test=170 (120+). External test106 (56+) | External; geographical split | Lesion sensitivity: 85.7%; PPV: 91.5%; Accuracy: 88.5%; Patient specificity: 98.0% |
| Yang et al 201126 | MRA-TOF | Retrospective single center | 287 cases; 92+cases; 147A; (1–31 mm); unruptured | 1 general and 1 neuroradiologist reviewed DSA | cCAD: Dot enhancement filter | Test=92 (92+) | No training required | Lesion sensitivity: 96%; FP/case: 11.6 |
| Timmins et al 202127 | MRA-TOF | Retrospective single center | 300 cases; 254+; 282A; 75% female; 55 years; 3.6 mm (1.0–15.9 mm); unruptured | 1 neuroradiologist and 1 trained reader reviewed MRA-TOF | DL: 3D CNN Retina U-net by MiBaumgartner | Training=113 cases (93+); Test=141 (115+) | Internal; hold out | Lesion sensitivity: 67%; FP/case: 0.13 |
| Nakao et al 201828 | MRA-TOF | Retrospective single center | 450 cases; 450+; 508A; 45% female; 61 years; 3 mm; unruptured | 2 radiologists reviewed MRA-TOF | DL: Voxel based CNN (chainer 1.6.1) | Training=300; Validation=50; Test=100 | Internal; temporal split | Lesion sensitivity: 94.2%; FP/case: 2.9 |
| Hanaoka et al 201929 | MRA-TOF | Retrospective single center | 300 cases; 300+; 300A | 2 radiologists reviewed MRA-TOF | cCAD: HoTPiG (voxel-based feature set) and Hessian based features with single SVM | Training=200; Test=100 | Internal; 3-fold CV | Lesion sensitivity: 80%; FP/case: 3 |
| Faron et al 202030 | MRA-TOF | Retrospective multicenter | 85 cases; 85+; 115A; 68% female; 56 years; 7.1 mm (2.1–37 mm); unruptured | 1 neuroradiologist reviewed MRA-TOF and other available imaging | DL: DeepMedic CNN | Training=58; Validation=10; Test=17 | Internal; CV (5-fold) | Lesion sensitivity: 90%; FP/case: 6.1 |
| Sichtermann et al 201931 | MRA-TOF | Retrospective multicenter | 85 cases; 85+; 115A; 68% female; 56 years; 7.1 mm (2.1–37 mm); unruptured | 2 radiologists reviewed MRA-TOF and other available imaging | DL: DeepMedic CNN | Training=58; Validation=10; Test=17 | Internal; CV (5-fold) | Lesion sensitivity: 90%; FP/case: 6.1 |
| Hou et al 202032 | MRA-TOF | Retrospective multicenter | 350 cases; 179+; 179A, 7.5 mm (2.4–23 mm); unruptured | NR | DL: 1D CNN by generating 1D vectors from MIP images | Training=245 (126+); Validation=35 (17+); Test=70 (36+) | Internal; hold out | Lesion sensitivity: 93.2%; Precision: 96.9%; Accuracy: 95.2%; AUC: 0.99;F1: 0.950 |
| Allenby et al 202133 | MRA-TOF | Retrospective multicenter | 623 cases; 21+; 21A; unruptured | 1 interventional radiologist reviewed MRA-TOF | cCAD: single-voxel morphometry | Test=623 (21+) | No training required | Patient specificity: 86%; Sensitivity: 81%; FP/case: 0.14 |
| Stember et al 201934 | MRA-TOF | Retrospective single center | 336 cases; 336+; 336A; (3–23 mm) | Original radiological reports | DL: U-net CNN | Training=250; Test=86 | Internal; hold out | Lesion sensitivity: 98.8%; AUC:0.87 |
| Chen et al 202035 | MRA-TOF | Retrospective single center | 131 cases; 131+; 140A, 63% female; 57 years; 6.5 mm; unruptured | 2 radiologists reviewed DSA | DL: 3D-U-net CNN | Training=76; Validation=20; Test=35 | Internal; hold out | Lesion sensitivity: 82.9%; FP/case: 0.86 |
| Ueda et al 201936 | MRA-TOF | Retrospective multicenter | 1271 cases; 1271+; 1477A; 72% female; 68 years; 4.1 mm | 2 radiologists reviewed MRA-TOF | DL: ResNet-18 using skip connections | Training and validation=683; Internal test=521; External test=67 | External; geographical split | Lesion sensitivity: 93%; FP/case: 5 |
| Nomura et al 202137 | MRA-TOF | Retrospective multicenter | 519 cases; 399+; 448A; 46% female; 3.1 mm | 2 radiologists reviewed MRA-TOF | DL/cCAD: 3 software: (i) 3D local intensity structure analysis; (ii) graph-based; (iii) CNN | Training=399 (339+); Test set 1=60 (30+); Test set 2=60 (30+) | External; geographical split | Per patient or per lesion performance measurements not available |
| Nemoto et al 201738 | MRA-TOF | Retrospective single center | 300 cases; 300+; 300A; 50% female; 59.8 years; 3.1 mm | 2 radiologists reviewed MRA-TOF | cCAD: voxel and candidate classifier ensembles with cost-sensitive AdaBoost | Training=200; Test=100 | Internal; 3-fold CV | Lesion sensitivity: 56.8%; FP/case: 10 |
| Hainc et al 202039 | 2D DSA | Retrospective single center | 240 cases; 136+; 186A; 65% female, 59 years, 7 mm; ruptured and unruptured | 2 neuro-interventional radiologists reviewed DSA | DL: commercial software by Cognex, ViDi Suite 2.0, Cognex Inc | Split via DSA Projections: 706 (335+) Training=565; Test=141 | Internal; 45-fold CV | Lesion sensitivity: 79%; Specificity: 79%;F1: 0.77; Precision: 0.75; Mean AUC: 0.76 |
| Zeng et al 202040 | 2D DSA | Retrospective single center | 300 cases; 250+; 263A | 5 radiologists reviewed DSA | DL: 2D CNN with spatial information fusion (VGG16) | No test set | No separate test set | No test set data |
| Jin et al 202041 | 2D DSA | Retrospective single center | 493 cases; 493+; 1205A; 62% female; 55 years; 7.4 mm (1.3–40 mm) | 2 neurologists reviewed DSA and third reader for arbitration | DL: end to end spatial temporal U-Net CNN (Keras-2.2.0 with TensorFlow-1.4.0 backend) | Training=249; Validation=98; Test=146 | Internal; hold out | Lesion sensitivity: 89.3%; Patient sensitivity: 97.7%; FP/case: 3.77 |
| Liu et al 202142 | 3D DSA | Retrospective single center | 451 cases; 451+; 485A; 61% female; 56 years; 7.1 mm | 2 neuroradiologists reviewed DSA | DL: 3D-Dense-Unet CNN | Training=347; Validation=41; Test=63 | Internal; hold out | Lesion sensitivity: 88.4%; FP/case: 0.61 |
| Hu et al 202043 | 3D DSA | Retrospective single center | 145 cases; 145+; 165A; 66% female; 57.8 years | 2 neuroradiologists reviewed DSA | cCAD: Bayesian optimized Hessian matrix filter | Test=145 | No training required | Lesion sensitivity: 96.4%; Precision: 0.946; AUC: 0.98; F1 score: 0.955 |
| Duan et al 201944 | 2D DSA | Prospective single center | 281 cases; 261+; 261A; 85% female; ruptured and unruptured | 2 radiologists reviewed DSA | DL: CNN based on feature pyramid networks (using ResNet50) | Training=241; Test=40 (20+) | Internal; temporal split | Patient sensitivity: 96.0%, Specificity: 91.0%; Accuracy: 93.5%; AUC: 0.94; F1 score: 0.94 |
| Rahmany et al 201845 | 2D DSA | Retrospective single center | 30 cases; 30+; 30A | 2 neuroradiologists reviewed DSA | cCAD: Priori knowledge applied to fuzzy logic-based model with Fuzzy information fusion | Test=30 | No training required | Results for 5 cases : Patient sensitivity: 100%; Specificity: 100%, Accuracy: 98.4%; AUC: 0.96 |
| Rahmany et al 201946 | 2D DSA | Retrospective single center | 30 cases; 30+; 30A | 2 neuroradiologists reviewed DSA | cCAD: LBP for feature extraction, and KNN classification | Training=20; Test=10 | Internal; hold out | Per patient or per lesion performance metrics not available |
| Malik et al 201847 | DSA | Retrospective single center | 59 cases; 47+; 47A; (6–21 mm); unruptured | 1 radiologist reviewed DSA | DL: classification multi-layer perceptron neural network | Split into 210 ROI: Training=189; Test=21 | Internal; 10-fold CV | Per patient or per lesion performance metrics not available |
| Chandra et al 201748 | 2D DSA | Retrospective single center | 15 cases; 15+; 15A | NR | cCAD: iterative double automated thresholding, morphological filtering | Test=15 | No training required | Per patient or per lesion performance metrics not available |
| Khan et al 201949 | DSA | Retrospective single center | 4 cases; 4+; 4A | NR | cCAD: sub-band morphological operation, gaussian filtering | Test=4 | No training required | Per patient or per lesion performance metrics not available |
| Shi et al 202050 | CTA | Retrospective multicenter | 1388 cases; 908+; 1145A; 31.3% female; 64 years; 4.4 mm; ruptured and unruptured | 3 neuroradiologists reviewed DSA | DL: end-to-end 3D CNN (DAResUNet) | Training=927 (744+); Validation=100 (50+); Testing: internal=150 (75+); external=211 (39+) | External; geographical split | Lesion sensitivity: 76.1%; PPV: 49.3%; NPV: 95.8%; Accuracy: 81.0%; Patient sensitivity: 84.6%; Specificity: 80.2%; FP/case: 0.27 |
| Shahzad et al 202051 | CTA | Retrospective single center | 253 cases; 253+; 294A; 67% female; 55.1 years; ruptured and unruptured | 1 neurosurgeon and 1 radiologist reviewed CTA and DSA (if available) | DL: 3D CNN based on DeepMedic | Training=68 (79+); Test=185 (215) | Internal; temporal split | Lesion sensitivity: 82%; F1: 0.66; Precision: 0.54; FP/case: 0.81 |
| Dai et al 202052 | CTA | Retrospective multicenter | 311 cases; 311+; 344A; 5.4 mm (1–24 mm) | 1 radiologist reviewed CTA | DL: RCNN model and Resnet-50 | Training=208; Test=103 | Internal; hold out | Lesion sensitivity: 91.8%; FP/case 8.9 |
| Hentschke et al 201453 | CE/TOF-MRA /CTA | Retrospective single center | 151 cases; 81+; 112A; (2.0–5.5 mm); CTA: 72; CE-MRA: 38; TOF-MRA: 41; unruptured | 2 neuroradiologists reviewed imaging | cCAD: sphere-enhancing filter and linear or non-linear classification | Test=151 (81+) | Internal; 4-fold CV | Lesion sensitivity: CE-MRA: 91%; TOF-MRA: 84%; CTA: 69%; FP/case: 10 |
| Lauric et al 201054 | 3D-RA/CTA | Retrospective single center | 20 cases; 19+; 20A; (3.2–10 mm); 3D-RA: 10; CTA:10 | 2 readers reviewed DSA/CT | cCAD: 3D shape analysis using writhe number | Test=20 (19+) | No training required | Lesion sensitivity: DSA: 100%; CTA: 100%; FP/case: DSA: 0.66; CTA: 5.36 |
+, aneurysm-positive cases; A, aneurysm; AUC, area-under-curve; cCAD, conventional computer assisted diagnosis; CE, contrast enhanced ; CNN, convolutional neural network; CTA, CT angiography; CV, cross-validation; 1D, one dimensional; 2D, two dimensional; 3D, three dimensional; DL, deep learning; 3D-RA, three dimensional rotational angiography; DSA, digital subtraction angiography; FP/case, false-positives per case; HoTPiG, histogram of triangular paths in graph; KNN, k-nearest neighbour; LBP, local binary patterns; LOO, leave-one-out; MIP, maximum intensity projection; ML, machine learning; MRA-TOF, MR angiography-time of flight; NPV, negative predictive value; NR, not recorded; PPV, positive predictive value; ROI, regions of interest; SVM, support vector machine.