Background
Motivation for left atrial fibrosis/scar segmentation challenge
State-of-the-art for cardiac fibrosis/scar segmentation
Reference | Model |
n
| Modality | LV/LA | Algorithm | Evaluation |
---|---|---|---|---|---|---|
Kim et al.[9] | Canine | 26 | CMR | LV | SD | Infarct size, ex-vivo |
Amado et al.[12] | Animal | 13 | CMR | LV | SD, FWHM | Bland altman, Infarct volume |
Kolipaka et al.[10] | Human | 23 | CMR | LV | SD | Percentage scar, Bland-Altman |
Positano et al.[14] | Human | 15 | CMR | LV | Clustering | Percentage scar |
Yan et al.[13] | Human | 144 | CMR | LV | SD | Percentage scar |
Schmidt et al.[11] | Human | 47 | CMR | LV | SD | Infarct size |
Hennemuth et al.[15] | Human | 21 | CMR | LV | EM fitting | Percentage scar, Bland-Altman |
Oakes et al.[5] | Human | 81 | CMR | LA | SD | Percentage scar |
Detsky et al.[16] | Human | 15 | CMR | LV | Clustering | Infarct size |
Tao et al.[17] | Human | 20 | CMR | LV | Otsu thresholding | Dice |
Knowles et al.[4] | Human | 7 | CMR | LA | MIP | Percentage scar |
Lu et al.[18] | Human | 10 | CMR | LV | Graph-cuts | Infarct size and Bland-Altman |
Proposed evaluation framework
Methods
Data acquisition database
U. Utah | BIDMC | KCL-IM | |
---|---|---|---|
Scanner type
| Siemens Avanto 1.5T or Vario 3T | Philips Acheiva 1.5T | Philips Achieva 1.5 T |
Basic params
| Free-breathing (FB) with navigator-gating | FB and navigator-gating with fat suppression | FB with navigator-gating with fat suppression |
TI
†
, TR, TE
| 300 ms, 5.4 ms, 2.3 ms | 280 ms, 5.3 ms, 2.1 ms | 280 ms, 5.3 ms, 2.1 ms |
Acquired resolution
| 1.25 × 1.25 × 2.5mm | 1.4 × 1.4 × 1.4mm | 1.3 × 1.3 × 4.0mm |
Pre-scan
| < 7 days | < 7 days | < 48 hours |
Post-scan
| 3 - 6 months | = 30 days | 3 - 6 months |
Algorithm | Technique | Evaluation | Atrial wall | Strengths | Weaknesses |
---|---|---|---|---|---|
IC: Bai et al. | Hysteresis thresholding | 30 pre and post | Euclidean distance - 3 mm | Coherent segmentations | Fixed sigmoid models derived from empirical data |
MV: Hennemuth et al. | Region-growing with EM-fitting | 30 pre and post | Euclidean distance - 3 mm | Post ablation imaging | Pre-ablation imaging |
SY: Lu et al. | MRF model with graph-cuts | 20 pre and post | Dilation - 4 mm | Fuzzy membership - improved delineation | Post-processing for small cluster removal |
HB: Gao et al. | Active contour and EM-fitting | 15 post | Active contour (snake) | Accurate myocardial segmentation | Fixed number of gaussian mixtures in model (i.e. two) |
YL: Peters et al. | Simple thresholding | 15 pre and post | Manual | Accurate segmentation on both pre- and post. | Time consuming |
KCL: Karim et al. | MRF model with graph-cuts | 30 pre and post | Post-ablation imaging | Pre-ablation imaging | Post-processing steps necessary |
UTA: Cates et al. | Histogram analysis and simple thresholding | 30 pre and post | Manual | Accurate segmentation on pre and post. | Time consuming |
UTB: Perry et al. | k-means clustering | 30 pre and post | Manual | Pre-ablation fibrosis | Equivalent variance across all clusters - LA scar variance more variable |
Algorithm 1: Imperial college - hysterisis thresholding (IC)
Background
Implementation
Algorithm 2: Mevis - Region growing with mixture model fitting (MV)
Background
Implementation
Algorithm 3: Sunnybrook - Graph-cuts with fuzzy c-means clustering (SY)
Background
Implementation
Algorithm 4: Harvard/Boston University - Active contours and mixture model fitting (HB)
Background
Implementation
Algorithm 5: Yale - Threshold selection with manual wall delineation (YL)
Background
Implementation
Algorithm 6: KCL - Graph-cuts with EM-algorithm (KCL)
Background
Implementation
Algorithm 7: Utah A - Threshold selection with manual wall delineation (UTA)
Background
Implementation
Algorithm 8: Utah B - Unsupervised learning using k-means clustering (UTB)
Background
Implementation
Algorithm evaluation
Reference standard 1: pseudo-ground truth
Reference standard 2: n-SD and FWHM
Evaluation metrics
Objective evaluation
Results
Segmentation accuracy with pseudo ground truth
Pre data | Post data | |||
---|---|---|---|---|
RMSE (mm) | |δV| (ml) | RMSE (mm) | |δV| (ml) | |
IC | 0.72 (0.5) | 2.87 (2.0) | 9.52 (8.2) | 4.79 (2.9) |
MV | 1.42 (0.7) | 38.08 (6.7) | 9.20 (8.8) | 4.15 (5.7) |
SY †∗ | 0.17 (0.1) | 12.87 (2.8) | 9.22 (9.3) | 10.19 (3.9) |
HB ∗ | n.a. | n.a. | n.a. | 20.16 (10.3) |
YL †∗ | 1.03 (0.4) | 0.62 (0.7) | 6.34 (8.2) | 2.77 (2.3) |
KCL | 1.33 (0.6) | 2.24 (2.2) | 9.20 (8.3) | 3.10 (2.3) |
UTA | 0.36 (0.3) | 3.24 (2.6) | 10.72 (8.0) | 3.54 (2.5) |
UTB | 0.52 (0.5) | 3.10 (2.2) | 8.91 (8.2) | 1.25 (1.5) |
2-SD | n.a. | 7.51 (3.6) | n.a. | 17.7 (10.1) |
3-SD | n.a. | 12.73 (8.3) | n.a. | 7.64 (3.7) |
4-SD | 0.15 (0.1) | 12.74 (8.3) | 11.69 (7.5) | 11.98 (8.5) |
6-SD | n.a. | n.a. | n.a. | 15.47 (8.5) |
FWHM | n.a. | 70.52 (38.4) | 7.67 (8.2) | 6.61 (5.9) |
Non-scar enhancing structures
Image quality on segmentation
Challengers | Good | Average | Poor |
---|---|---|---|
Mean (SD) | Mean (SD) | Mean (SD) | |
IC | 64 (26) | 64 (27) | 69 (23) |
MV | 83 (20) | 80 (21) | 79 (20) |
SY* | 70 (21) | 64 (26) | 71 (22) |
HB* | 76 (17) | 71 (21) | 74 (16) |
YL* | 80 (20) | 73 (25) | 74 (24) |
KCL | 78 (18) | 77 (25) | 73 (26) |
UTA | 64 (29) | 70 (29) | 71 (28) |
UTB | 63 (28) | 67 (28) | 67 (24) |
2-SD | 56 (27) | 53 (29) | 53 (27) |
4-SD | 17 (19) | 21 (27) | 17 (15) |
6-SD | 38 (21) | 35 (25) | 34 (20) |
FWHM | 68 (30) | 66 (27) | 56 (34) |
Discussion
Evaluation framework
Evaluated algorithms
Method | Normalisation | Model |
---|---|---|
IC | Y | Sigmoid |
MV | Y | Gaussian |
SY | N | Gaussian |
HB | N | Gaussian |
YL | N | None |
KCL | Y | Gaussian |
UTA | N | Gaussian |
UTB | Y | Gaussian |