<aside> 💡 Summary : A project on abdominal vessel segmentation using CT scans, particularly in the context of gastrectomy procedures. The study made use of around 300 pairs of CT scans procured from Hutom Inc., each of which included both arterial and portal phases.
The project utilized a MONAI-based segmentation framework to train a 3D Dense UNet model. The initial results showed a dice coefficient of 0.8 and a sensitivity of 0.7. However, due to the resampling process, the distal vasculature got destroyed, causing significant disconnections in the vessels.
Further investigations and experiments led to a set of results where, despite a decrease in the dice score due to thicker estimates, there was a marked improvement in sensitivity. Additionally, no further false positives were observed.
Expert qualitative analysis revealed that the model, when trained with morphological preprocessing on the Ground Truth (GT), provided clear and accurate vessel predictions. The study underscores the importance of proper image processing in effective training. It also suggests the potential need for a different metric to accurately demonstrate the quality of vascular segmentation and disconnection.
</aside>
Gastrectomy is a common surgical procedure in South Korea. Multiple clinical studies have shown the impact of blood loss during the procedure [1]. With the development of medical imaging AI and the increasing importance of medical twin technology, I studied abdominal vessel segmentation on computed tomography (CT).
morphologicalTwo different CT phases provide distinct vessel information. The arterial phase (AP) displays arteries with highly enhanced HU values, while the portal phase (PP), captured immediately after the arterial phase on the same patient, shows enhanced veins.
I obtained around 300 CT pairs from Hutom Inc., each including both arterial and portal phases. As part of preprocessing, I resampled the pixel spacing, adjusted the slice thickness, and discarded scout images. The input images were treated with special attention and exception methods during preprocessing to ensure effective learning, with a particular focus on the vessel structure.
I used the MONAI-based segmentation framework to train multiple 3D segmentation models such as Dense UNet (with-/without- Dilation) and the new models that I designed the architectures. Additionally, I optimized the final model based on the test data.
I tested multiple augmentation methods like random HU windowing, mask-based removals and so on.
To optimize the vessel inferences, well-designed loss function is essential. Because the target classes has long and thin structures. The new loss function is designed to focus on vessel’s stream rather than each pixel like dice.
According to specialists' qualitative analysis, the models have evolved continuously. Compared to nnUNet, the first model I released has significantly better sensitivity. However, it creates clustered vessel bunches when they are closely located because the estimated vessels are too thick to distinguish each line. The second model, with a new loss function and a new architecture of the model, addresses these issues. We have observed thinner vessels with less loss in sensitivity.
Abdominal Artery (except Aorta) Segmentation Performance in Dice score for each threshold values
Abdominal Artery (except Aorta) Segmentation Performance in Recall score for each threshold values