Deep Learning powered approaches are the current state-of-the-art method for segmenting Coronary Arteries. One of the biggest challenges is the time-intense reliance on a human expert for annotations.
Segmenting coronary arteries is challenging due to their small size and winding anatomy leading to low-resolution and artifacts. However, automatic segmentation is not equally hard for all parts of the vessel. In fact, most proximal parts are rather homogenous and can easily be segmented by a neural network. In contrast, many critical areas, such as branches, plaques or other diseased areas, can be very diverse and challenging for deep-learning models. The heterogeneous nature of the task leads us to hypothesize that it might be well suited for an active-learning (AL) approach (Figure 1). Implementing an Active Learning Pipeline allows us to focus on the critical areas, in which segmentation often fails, drastically cutting down on expensive, human-expert annotation time.
Methods:
As a pre-processing step the centerlines of the coronary arteries were automatically extracted and used to create curved planar reformations (CPR) of the vessels. Using a 3D-CNN on stacks of axial images, we trained a model as a baseline with data from 5 randomly selected patients, a total of 57 CPRs and 33501 slices of 2D-data.
We implemented an Active Learning cycle by using the trained baseline network to predict on unlabeled data. Mislabeled data is then selected and corrected by a human-in-the-loop. The corrected data is added to the training-base and the cycle is repeated (Figure 1). Because only faulty areas need to be manually corrected this method of training is very annotation-efficient. The AL-cycle was repeated twice, first with three and afterwards with two patients.
We compared the resulting model with the baseline on a withheld test set of 5 patients. Comparison metrics were the Dice-score (DSC) and Hausdorff Distance 95% Percentile (HD95), two well established evaluation metrics for this task. In order to highlight the performance-difference for crucial areas, part of the comparison focused on branches and more challenging segmentation-tasks, excluding areas with uncertain ground-truth (i.e. due to artifacts) (Figure 2).
Results:
Compared to annotating all the vessels used for Active Learning, we only needed to manually correct 23% of all slices. As a consequence total human annotation time was cut down by more than 65%. Though the necessary annotation was comparatively small, our resulting model outperformed the baseline-model significantly (DSC: 0.807 vs 0.727, P<.001 | HD95: 3.75 vs. 7.16 P<.001).Comparing the performance on pre-selected branches and error-prone areas, our resulting model outperformed the baseline by (Dice: 0.820 vs 0.694, P<.001 | HD95: 9.2 vs. 14.02 P<.001).
Conclusion:
We successfully showcased the use of Active Learning for Coronary Artery Segmentation with a simple, reproducible pipeline. Furthermore we show that even with a small amount of corrections, training can be improved significantly for critical areas.
Our setup allows us to focus the expensive resource of expert-annotations on where they are most effective, allowing for more training-data and quicker performance gains. We argue that together with automated uncertainty-detection methods this framework can be used to improve efficiency and stability of artificial intelligence for coronary artery segmentation.
https://dgk.org/kongress_programme/ht2022/aP350.html