Chronic pancreatitis (CP) is a long-standing inflammatory condition that can lead to irreversible damage to the pancreas. This condition is often accompanied by complications such as persistent abdominal pain, pancreatic insufficiency, diabetes and a heightened risk of cancer. Accurate imaging is essential for diagnosis, monitoring and treatment planning. Computed tomography (CT) scans are widely used in clinical settings, particularly in the portal venous phase, to assess the pancreas. However, manual segmentation of the pancreas from CT scans is time-consuming and subject to inter-reader variability. Moreover, the pancreas has a highly variable anatomy, which becomes even more complex in CP due to fibrosis, calcification, atrophy and other structural changes.
Existing automated segmentation models based on artificial intelligence (AI) have shown promise, but most are trained on healthy individuals or patients with pancreatic tumours. These models often fail when applied to CP patients because of the significant morphological differences. To address this limitation, researchers developed a new AI-based segmentation model trained on a large, mixed dataset that includes both healthy subjects and individuals with CP. The aim was to build a model capable of accurately segmenting the pancreas across a broad spectrum of anatomical variations and imaging conditions.
Training a Model for Clinical Complexity
The AI model was developed using the nnU-Net architecture, a self-configuring deep learning framework known for its versatility in medical image segmentation. The training dataset included 326 CT scans from Aalborg University Hospital in Denmark, comprising both CP patients and healthy controls. These scans represented a wide range of anatomical and pathological presentations. The model was then tested internally on 47 cases from the same hospital and externally on datasets from Bergen Hospital in Norway and the National Institutes of Health (NIH) in the United States.
Must Read: Improving Pancreatic Cyst Follow-Up with Multimodal AI
Performance was evaluated using the Dice score, a standard metric for measuring segmentation accuracy. The model achieved a high Dice score of 0.85 on the internal test set, and consistent scores of 0.79 on both external datasets. These results highlight the model’s robustness and adaptability across different patient populations, scanner models and imaging protocols. Notably, this performance is comparable to, or better than, other leading AI segmentation models trained solely on healthy subjects.
Ground truth labels for training were meticulously prepared by junior radiologists under strict guidelines, with oversight from an experienced radiologist when necessary. This ensured the reliability of the annotations, which is critical for effective AI training. Importantly, the inclusion of data from various scanner types and imaging protocols enhanced the model’s generalisability, making it suitable for real-world clinical use.
Understanding Model Performance
A key finding of the study was that anatomical and physiological factors had a greater influence on segmentation accuracy than technical CT parameters. Higher visceral fat and larger pancreas volume were associated with better model performance. This is likely because increased visceral fat provides better contrast between the pancreas and surrounding tissue, facilitating clearer delineation. Conversely, patients with low visceral fat or smaller pancreas size posed more challenges, leading to lower Dice scores.
In some extreme cases, particularly those involving heavy fat infiltration of the pancreas, the model’s performance dropped significantly, with Dice scores as low as 0.00. However, such cases were rare and did not substantially affect the overall performance metrics. The external NIH dataset also presented challenges due to differences in segmentation protocols and a higher prevalence of low visceral fat among subjects. These discrepancies emphasise the need for standardised labelling guidelines in AI research to enable more accurate benchmarking and cross-study comparisons.
Interestingly, variations in scanner models, slice thickness, contrast agents and other CT parameters did not significantly affect the model’s performance. This finding is encouraging for clinical implementation, as it suggests that the model can be deployed across various healthcare settings without requiring stringent standardisation of imaging protocols.
Clinical Relevance and Future Directions
This AI segmentation model holds significant promise for both clinical practice and research. By automating the process of pancreas segmentation, it reduces the workload on radiologists and ensures more consistent and reproducible results. These improvements are particularly relevant in the context of CP, where accurate imaging is essential for tracking disease progression, evaluating treatment efficacy and identifying early structural changes.
Looking ahead, the integration of this segmentation model into a broader AI-based diagnostic pipeline could further enhance clinical decision-making. Combining segmentation with radiomic feature extraction, disease classification and predictive analytics could lead to the identification of novel imaging biomarkers. These tools could facilitate earlier diagnosis and more precise monitoring of CP, potentially improving patient outcomes.
The study also points to opportunities for future refinement. For instance, incorporating more data from cases with severe fat infiltration could help the model learn to manage these difficult scenarios. Exploring alternative architectures, such as Transformer-based models, may offer additional improvements in handling complex anatomical variations. Furthermore, enhancing the interpretability of the model through visualisation techniques could aid clinicians in understanding and trusting AI-generated outputs.
The development of an AI-based pancreas segmentation model trained on both healthy individuals and CP patients represents a significant advance in medical imaging. With strong performance across varied datasets and minimal sensitivity to technical parameters, the model demonstrates a high degree of robustness. Its deployment in clinical and research settings could streamline pancreas imaging, support the identification of novel biomarkers and ultimately improve the care of patients with chronic pancreatitis.
Source: European Journal of Radiology
Image Credit: iStock
