Automatic Segmentation of the Left Atrium from Computed Tomography Angiography Images.

The left atrial appendage (LAA) causes 91% of thrombi in atrial fibrillation patients, a potential harbinger of stroke. Leveraging computed tomography angiography (CTA) images, radiologists interpret the left atrium (LA) and LAA geometries to stratify stroke risk. Nevertheless, accurate LA segmentation remains a time-consuming task with high inter-observer variability. Binary masks of the LA and their corresponding CTA images were used to train and test a 3D U-Net to automate LA segmentation. One model was trained using the entire unified-image-volume while a second model was trained on regional patch-volumes which were run for inference and then assimilated back into the full volume. The unified-image-volume U-Net achieved median DSCs of 0.92 and 0.88 for the train and test sets, respectively; the patch-volume U-Net achieved median DSCs of 0.90 and 0.89 for the train and test sets, respectively. This indicates that the unified-image-volume and patch-volume U-Net models captured up to 88 and 89% of the LA/LAA boundary's regional complexity, respectively. Additionally, the results indicate that the LA/LAA were fully captured in most of the predicted segmentations. By automating the segmentation process, our deep learning model can expedite LA/LAA shape, informing stratification of stroke risk.

Convolutional Neural Network based Segmentation of Abdominal Aortic Aneurysms.

Abdominal aortic aneurysms (AAAs) are balloonlike dilations in the descending aorta associated with high mortality rates. Between 2009 and 2019, reported ruptured AAAs resulted in ~28,000 deaths while reported unruptured AAAs led to ~15,000 deaths. Automating identification of the presence, 3D geometric structure, and precise location of AAAs can inform clinical risk of AAA rupture and timely interventions. We investigate the feasibility of automatic segmentation of AAAs, inclusive of the aorta, aneurysm sac, intra-luminal thrombus, and surrounding calcifications, using 30 patient-specific computed tomography angiograms (CTAs). Binary masks of the AAA and their corresponding CTA images were used to train and test a 3D U-Net - a convolutional neural network (CNN) - model to automate AAA detection. We also studied model-specific convergence and overall segmentation accuracy via a loss-function developed based on the Dice Similarity Coefficient (DSC) for overlap between the predicted and actual segmentation masks. Further, we determined optimum probability thresholds (OPTs) for voxel-level probability outputs of a given model to optimize the DSC in our training set, and utilized 3D volume rendering with the visualization tool kit (VTK) to validate the same and inform the parameter optimization exercise. We examined model-specific consistency with regard to improving accuracy by training the CNN with incrementally increasing training samples and examining trends in DSC and corresponding OPTs that determine AAA segmentations. Our final trained models consistently produced automatic segmentations that were visually accurate with train and test set losses in inference converging as our training sample size increased. Transfer learning led to improvements in DSC loss in inference, with the median OPT of both the training segmentations and testing segmentations approaching 0.5, as more training samples were utilized.