Catheter Ablation -> Atrial Fibrillation & Atrial Flutter: -> Mapping & Imaging D-PO01 - Featured Poster Session (ID 11) Poster

D-PO01-145 - Utilizing A Multi-label Convolutional Neural Network Removes Inter-observer Variability In Assessment Of Atrial Fibrosis (ID 111)


Background: Late gadolinium enhancement (LGE) scans provide non-invasive estimate of atrial fibrosis. However, their widespread adoption has been hindered partly by minimal validation and non-standardized image processing techniques, which are operator and algorithm dependent.
Objective: To quantify atrial fibrosis from LGE scans using an open source, operator independent, automatic pipeline.
Methods: A multi-label convolutional neural network (Fig. A) was designed to delineate atrial structures including the blood pool, pulmonary veins and mitral valve. The network removed the operator dependent segmentation step and allowed for implementation of a fibrosis quantification pipeline (Fig. B). The results (Fig. C) were compared against manual global fibrosis burdens, calculated using published thresholding techniques: image intensity ratio (IIR) 0.97, IIR 1.61, and mean of blood pool signal intensity +3.3SD.
Results: The pipeline was validated on a large LGE dataset (n=207), manually analyzed by 5 operators. Segmentations by the network achieved a 91% Dice score against the ground truth, in contrast to an 85% score in the inter-observer analysis. Intra-class correlation coefficients (ICC) of automatic fibrosis burdens with the ground truth were superior to inter-observer correlations for all 3 thresholds: IIR 0.97 (0.94 vs 0.88), IIR 1.61 (0.99 vs 0.99), and +3.3SD (0.99 vs 0.96). The analysis required 3 minutes per case on a standard desktop.
Conclusion: This pipeline provides an automatic fibrosis quantification method that is superior to manual analysis and removes image processing variability in assessment of LGE scans, which can help with their wider clinical adoption.