Accurate organ segmentation in CT images is a foundational step for computer-aided diagnosis, surgical planning, and treatment response monitoring. Most organs such as the liver or spleen are large and have distinct boundaries, but the pancreas presents exceptional challenges for automated segmentation algorithms.
The pancreas is a small organ with a highly variable shape across patients, ranging from tadpole-like to elongated. It has low contrast with surrounding fat and connective tissues on CT scans, and its boundaries are often difficult to distinguish even for experienced radiologists.
When patients have pancreatic diseases such as pancreatic ductal adenocarcinoma (PDAC) or pancreatitis, the appearance of the organ is further distorted, making reliable automated segmentation even more difficult. These challenges motivated the development of a specialized deep learning approach for diseased pancreata.
This study builds upon the Richer Convolutional Features (RCF) network, originally developed for edge detection in natural images. The RCF architecture extracts feature maps at multiple convolutional layers simultaneously, capturing both fine-grained local details and broader contextual information.
The key innovation proposed in this paper is a multi-layer upsampling extension that aggregates feature maps from several different network layers rather than relying on a single final prediction. Each intermediate feature map is upsampled to the original image resolution and combined, allowing the model to leverage complementary information from both shallow and deep network layers.
This hierarchical aggregation strategy is particularly valuable for pancreas segmentation because the organ requires both high-resolution boundary detail (captured by shallow layers) and semantic understanding of anatomical context (captured by deeper layers) to delineate accurately.
The study used a dataset of 59 patients, including 44 patients with pancreatic diseases such as PDAC, pancreatitis, and cysts, plus 15 healthy controls. A total of 236 CT slices were included after selection, with each slice manually annotated by radiologists to serve as ground truth segmentation.
Data was split into training and testing sets, and data augmentation techniques including rotation, flipping, and intensity jitter were applied to artificially expand the limited training set and improve model generalizability to unseen patients.
The model was trained using a weighted cross-entropy loss function that accounts for the significant class imbalance between pancreas pixels (which represent a small fraction of the CT slice) and background pixels. Without this weighting, the model would tend to predict background everywhere and still achieve high pixel accuracy.
Both single-phase (portal venous) and multi-phase CT inputs were evaluated. Multi-phase CT provides images acquired at different time points after contrast injection, capturing how tissues enhance over time and providing additional discriminative information beyond a single acquisition.
The proposed multi-layer upsampling model achieved a Dice Similarity Coefficient (DSC) of 76.36% on the test set when using multi-scale image inputs. This represents a substantial improvement over the single-scale baseline and demonstrates the benefit of aggregating predictions across multiple network layers.
Performance varied across disease subgroups. Segmentation was generally more accurate for normal pancreata compared to diseased cases, reflecting the greater shape variability and altered tissue appearance in patients with PDAC or pancreatitis.
The model showed particular improvement for multi-phase CT inputs compared to single-phase acquisitions. Incorporating information from multiple contrast phases allowed the network to leverage temporal enhancement patterns as an additional discriminative signal for delineating pancreatic boundaries.
A key limitation of this study is the relatively small dataset of 59 patients. Deep learning models generally benefit from larger training sets to learn robust feature representations, and performance on this dataset may not reflect results on larger, more diverse patient populations across different imaging centers.
The study focused on 2D segmentation of individual CT slices rather than full 3D volumetric segmentation of the entire pancreas. While 2D approaches are computationally simpler, 3D methods that exploit inter-slice continuity may achieve more anatomically consistent results for clinical applications.
Accurate automated pancreas segmentation has significant clinical implications. It can serve as a preprocessing step for downstream tumor detection and characterization, enabling fully automated pipelines that reduce radiologist workload and provide consistent, reproducible measurements of tumor size and response to therapy.