Introduction
The pancreas presents a significant challenge for automated segmentation in CT imaging. Unlike organs with relatively stable morphology, such as the liver or kidneys, the pancreas is highly variable in shape and position, often influenced by the deformation of adjacent structures. Traditional atlas-based methods frequently struggle with this variability. This paper proposes a hybrid framework that leverages the semantic strengths of 3D Fully Convolutional Networks (FCNs) and the spatial regression capabilities of Random Forests to achieve robust localization and segmentation.
Methodology
Localization via Regression Forests
The first stage of the pipeline involves identifying the pancreas bounding box. The authors employ a Regression Forest that utilizes a diverse set of features, including intensity differences and organ likelihood maps. A key innovation is the use of 3D FCN features extracted from a network trained on multiple abdominal organs. These features provide critical anatomical context, allowing the model to "triangulate" the pancreas based on the positions of more stable organs like the liver or spine.
LBP-like Encoding of Deep Features
To bridge the gap between high-dimensional deep learning features and the voxel-level processing of regression forests, the authors introduced a Local Binary Pattern (LBP)-like encoding. This method extracts 64-dimensional feature vectors from the penultimate layer of a 3D U-Net and converts them into a texture descriptor. By analyzing three orthogonal planes (axial, coronal, and sagittal) passing through a patch center, the system captures the statistical signature of the anatomy, making localization more robust to noise and contrast variations.
Segmentation Refinement
Following localization, a patient-specific probabilistic atlas is generated within the predicted bounding box. This atlas is then refined using a Graph Cut algorithm to produce the final volumetric segmentation of the pancreas.
Results and Discussion
The proposed method was evaluated on 146 abdominal CT volumes. The inclusion of 3D FCN features improved localization accuracy, reducing the average distance error for bounding box faces from 11.4 mm to 11.0 mm. The final segmentation achieved a Dice Overlap score of 73.9%.
Notably, the hybrid approach showed significant gains in difficult cases. In instances where previous methods failed (achieving Dice scores below 30%), the new feature set allowed for dramatic improvements. For example, in one challenging case, the Dice score jumped from 28.0% to 72.7%. This suggests that deep-learned contextual features are highly effective at preventing the model from being misled by adjacent organs like the stomach.
Conclusion
By combining 3D FCNs with regression forests, this research demonstrates that high-level semantic information can significantly enhance the performance of traditional machine learning models in complex medical imaging tasks. While the multi-step pipeline is computationally more intensive than some end-to-end networks, this hybrid approach offers superior robustness in anatomically challenging scenarios.
Summary updated