A Bottom-Up Approach for Automatic Pancreas Segmentation in Abdominal CT Scans

Organ segmentation is a prerequisite for a computer-aided diagnosis (CAD) system to detect pathologies and perform quantitative analysis. For anatomically high-variability abdominal organs such as the pancreas, previous segmentation works report low accuracies when comparing to organs like the heart or liver. In this paper, a fully-automated bottom-up method is presented for pancreas segmentation, using abdominal computed tomography (CT) scans. The method is based on a hierarchical two-tiered information propagation by classifying image patches. It labels superpixels as pancreas or not via pooling patch-level confidences on 2D CT slices over-segmented by the Simple Linear Iterative Clustering approach. A supervised random forest (RF) classifier is trained on the patch level and a two-level cascade of RFs is applied at the superpixel level, coupled with multi-channel feature extraction, respectively. On six-fold cross-validation using 80 patient CT volumes, we achieved 68.8% Dice coefficient and 57.2% Jaccard Index, comparable to or slightly better than published state-of-the-art methods.

💡 Research Summary

This paper tackles the challenging problem of automatically segmenting the pancreas in abdominal computed tomography (CT) scans, where high anatomical variability and low contrast make the organ difficult to delineate. The authors propose a fully automated, bottom‑up framework that relies on hierarchical, two‑tiered information propagation rather than global shape priors.

First, each axial CT slice is over‑segmented using Simple Linear Iterative Clustering (SLIC), producing a few hundred super‑pixels per slice. Within each super‑pixel, dense image patches (e.g., 16 × 16 px) are extracted. For every patch, a three‑channel representation is built (original CT intensity, gradient magnitude, and Laplacian) and a set of 25 statistical descriptors is computed, including mean, variance, entropy, and several Gray‑Level Co‑occurrence Matrix (GLCM) texture measures. These patch‑level features train a Random Forest (RF) classifier that outputs a probability (confidence) that the patch belongs to the pancreas.

The patch confidences are pooled across all patches belonging to the same super‑pixel, yielding an average confidence that becomes part of the super‑pixel’s feature vector. At the super‑pixel level, a cascade of two RF classifiers is applied. The first RF combines the pooled confidence with geometric, intensity, and texture descriptors of the super‑pixel to perform a coarse binary decision (pancreas vs. background). Super‑pixels classified as positive are passed to a second RF that refines the decision using additional multi‑channel features (intensity histograms, spatial relationships to neighboring super‑pixels) and higher‑order statistics of the patch‑level confidence distribution (variance, max, skewness). This hierarchical approach allows local evidence from small patches to be aggregated into a robust, region‑level decision without imposing a global shape model.

The method was evaluated on a dataset of 80 abdominal CT volumes, each containing roughly 200–300 axial slices. A six‑fold cross‑validation scheme was employed. The primary performance metrics were the Dice similarity coefficient and the Jaccard index. The proposed system achieved an average Dice of 68.8 % and a Jaccard of 57.2 %. These results are comparable to, and in some cases slightly better than, recent state‑of‑the‑art deep‑learning approaches (e.g., 3‑D U‑Net variants reporting Dice scores in the 65–70 % range) and substantially outperform traditional shape‑based or atlas‑based methods that typically score below 55 % Dice for the pancreas.

Key strengths of the approach include: (1) independence from global anatomical priors, which makes it tolerant to the pancreas’s large shape variations; (2) the interpretability and relatively low computational cost of Random Forest models; (3) effective exploitation of multi‑scale information by merging fine‑grained patch confidences with super‑pixel context. However, the study also has limitations. Processing is performed slice‑by‑slice, so inter‑slice continuity is not explicitly modeled, potentially leading to small gaps or fragmented regions. Random Forests, while efficient, may not capture the highly non‑linear patterns that deep convolutional networks can learn, especially in low‑contrast or noisy CT data. Finally, the dataset size (80 patients) is modest, leaving open questions about generalization to larger, more diverse populations.

Future work suggested by the authors includes extending the framework to three‑dimensional context (e.g., 3‑D super‑voxels and 3‑D RF cascades), integrating convolutional neural networks to provide richer feature embeddings while retaining the hierarchical decision structure, and leveraging multimodal imaging (MRI, PET) or transfer learning to improve robustness. Building larger, publicly available pancreas segmentation benchmarks would also enable more rigorous comparisons between bottom‑up, top‑down, and hybrid strategies.

In summary, the paper presents a novel, bottom‑up pancreas segmentation pipeline that achieves competitive accuracy through a clever combination of patch‑level Random Forest confidences and a two‑stage super‑pixel classifier. Its design highlights the potential of hierarchical, data‑driven methods for organs with high anatomical variability, and it opens avenues for further enhancements using 3‑D information and deep learning techniques.