Leveraging Anatomical Priors for Automated Pancreas Segmentation on Abdominal CT

Abstract: An accurate segmentation of the pancreas on CT is crucial to identify pancreatic pathologies and extract imaging-based biomarkers. However, prior research on pancreas segmentation has primarily focused on modifying the segmentation model architecture or utilizing pre- and post-processing techniques. In this article, we investigate the utility of anatomical priors to enhance the segmentation performance of the pancreas. Two 3D full-resolution nnU-Net models were trained, one with 8 refined labels from the public PANORAMA dataset, and another that combined them with labels derived from the public TotalSegmentator (TS) tool. The addition of anatomical priors resulted in a 6\% increase in Dice score ($p < .001$) and a 36.5 mm decrease in Hausdorff distance for pancreas segmentation ($p < .001$). Moreover, the pancreas was always detected when anatomy priors were used, whereas there were 8 instances of failed detections without their use. The use of anatomy priors shows promise for pancreas segmentation and subsequent derivation of imaging biomarkers.

Leveraging Anatomical Priors for Automated Pancreas Segmentation on Abdominal CT

The paper "Leveraging Anatomical Priors for Automated Pancreas Segmentation on Abdominal CT" presents an innovative approach to improving the segmentation performance of the pancreas on computed tomography (CT) images. Unlike previous methods that primarily focus on modifying segmentation model architectures or applying pre- and post-processing techniques, this study explores the application of anatomical priors to guide segmentation models effectively.

Methodology Overview

The researchers employed two 3D full-resolution nnU-Net models: one trained with refined labels from the public PANORAMA dataset and another that included labels derived from the public TotalSegmentator (TS) tool. The addition of anatomical priors facilitated a significant improvement in the Dice score, with an increase of 6% ($p < .001$) and a noticeable reduction in the Hausdorff distance error by 36.5 mm ($p < .001$). This advancement demonstrates the utility of anatomical priors in differentiating the pancreas from neighboring anatomical structures, reducing failed detections considerably when these priors are used.

The study utilized patient data from two public datasets containing abdominal CT scans—PANORAMA and AMOS22. PANORAMA contributed 1,350 portal-venous CT scans for model training, while AMOS22 provided 500 contrast-enhanced CT scans for external evaluation. The implementation involved dividing the labeled pancreatic arteries into aorta and remaining arteries to maintain consistency in evaluations compared to AMOS22 labels.

Results and Discussion

A quantitative evaluation revealed that the model incorporating anatomical priors (ALL$\_$45) achieved superior pancreas segmentation performance compared to the model without these priors (REF$\_$8). The ALL$\_$45 model demonstrated a Dice score of 0.81 $\pm$ 0.14, surpassing the REF$\_$8 score of 0.75 $\pm$ 0.22. Importantly, ALL$\_$45 ensured pancreas detection in every test instance with no failures, unlike the REF$\_$8 model, which recorded eight missed detections.

These results indicate the crucial role anatomical priors play in constraining segmentation models for more accurate delineation of the pancreas, which is otherwise challenging due to its proximity to structures of similar intensity and variability in morphology across different CT scans. Undoubtedly, this methodology offers a robust guideline for enhancing segmentation tasks in medical imaging, leveraging the spatial relationships of anatomical structures to refine model outputs.

Implications and Future Directions

This research highlights the potential for automated extraction of imaging biomarkers crucial for early diagnosis of pancreatic diseases, such as pancreatic cancer and diabetes. The findings advocate for wider adoption of anatomical priors in segmentation tasks to improve accuracy and reliability, especially in complex anatomical regions like the abdomen.

While the study specifically addresses contrast-enhanced CT datasets, the exploration of these models' efficacy across non-contrast CT images remains a prospective area for future investigation. Additionally, the methodology could expand to encompass other organs that demand precise segmentation due to their diagnostic significance.

In conclusion, this paper provides valuable insights into the role of anatomical priors in advancing segmentation technology and sets a precedent for future research in leveraging these constraints within automated deep learning pipelines for medical imaging.