Introduction
Pancreas segmentation in Computed Tomography (CT) scans is a critical yet challenging task in medical image analysis. The difficulty stems from three primary factors: the organ's small volume—typically occupying less than 0.5% of an abdominal scan—its high inter-patient variability in shape and location, and low-contrast boundaries that often blend with adjacent structures like the stomach, duodenum, and intestines. While two-stage deep learning models have improved localization, they often struggle to capture the complex, thin, and "wiggly" geometry of the pancreas.
Methodology
This paper proposes a two-stage cascaded framework designed to enhance geometric understanding through multitask learning.
Cascaded Framework
- Stage 1 ($C_1$): Coarse Localization. A 3-D Fully Convolutional Network (FCN) processes downsampled CT images to identify a bounding box around the pancreas, effectively filtering out irrelevant abdominal noise.
- Stage 2 ($C_2$): Fine Multitask Segmentation. Within the localized region, a 3-D Dense-U-Net performs two simultaneous tasks: voxel-wise segmentation (predicting pancreas vs. background) and skeleton extraction (a regression task predicting the organ's internal skeleton).
Multitask Learning and Skeletonization
The core innovation lies in treating segmentation and skeletonization as complementary tasks. While segmentation focuses on boundary detection, skeletonization captures the organ's internal topology. This dual focus forces the model to develop a robust internal representation of the pancreas's global shape, which helps maintain accuracy when boundaries are blurry.
The architecture incorporates dense connections to facilitate feature flow and a "Feature Consistency Module" to fuse low-level and high-level features effectively. Finally, a 3-D Fully Connected Conditional Random Field (CRF) is applied as a post-processing step to smooth labels and eliminate isolated false-positive clusters.
Experimental Results
The model was evaluated on the NIH Pancreas Dataset (82 scans) and an in-house dataset from Fujian Medical University (FMU, 59 scans) containing complex pathological cases with benign or malignant cysts.
Performance Metrics
On the NIH benchmark, the proposed method achieved a state-of-the-art average Dice Similarity Coefficient (DSC) of 86.4%, surpassing previous models like RSTN (84.5%). The model also demonstrated significant robustness on the FMU dataset, handling morphological deformations caused by pathology more effectively than standard networks.
Efficiency and Ablation
The localization stage ($C_1$) is highly efficient, requiring approximately 4 seconds on an NVIDIA TITAN XP GPU. Ablation studies confirmed that both skeleton guidance and dense connections are essential for achieving peak accuracy; removing these components led to a measurable drop in performance.
Discussion and Conclusion
The integration of skeletonization as a geometric prior provides a significant advantage in segmenting organs with high shape variability. By providing a structural "reference," the multitask approach prevents the model from under-estimating the organ volume when boundaries are indistinct. Future work may extend this "segmentation + skeleton" paradigm to other tubular or thin structures, such as blood vessels or the gastrointestinal tract. While the model is robust, extremely large tumors that completely disrupt the underlying skeleton remain a challenge, suggesting further refinement for advanced pathological cases.
Summary updated