A Unified Framework for the Detection and Classification of Fatty Pancreas in Ultrasound Images

Technical Summary: A Unified Framework for the Detection and Classification of Fatty Pancreas in Ultrasound Images

Problem Statement
Non-alcoholic fatty pancreas disease (NAFPD) is a significant medical condition linked to metabolic syndrome, insulin resistance, and an increased risk of pancreatic cancer. However, its diagnosis remains challenging due to the subjective nature of conventional ultrasound evaluation, high inter-observer variability, and the anatomical difficulty of visualizing the pancreas. While advanced modalities like CT and MRI offer objective quantification, they are costly, involve radiation (CT), or lack universal availability. Ultrasound remains the primary, cost-effective screening tool, but current assessment relies heavily on manual visual grading of echogenicity, leading to inconsistent prevalence reports and severity classifications. There is a lack of fully automated pipelines capable of classifying fatty pancreas from ultrasound images.

Methodology
The authors propose an end-to-end framework that mimics the clinical reasoning process of radiologists, consisting of three distinct stages:

1. Segmentation:

   • Architecture: The system employs a TransUNet-based architecture, combining a ResNet50 encoder with a transformer bottleneck (utilizing multi-head self-attention) and a dilated convolution module. This design captures both local texture details and global context.
   • Targets: Two separate models are trained to segment the pancreatic parenchyma and the splenic vein.
   • Transfer Learning: To address data scarcity and overfitting, the models are initialized via domain-specific transfer learning from a liver segmentation task (sharing similar ultrasound characteristics) before being fine-tuned on the target pancreas/splenic vein dataset.
   • Loss Function: Training utilizes a combined Dice and Binary Cross-Entropy (BCE) loss.

2. Anatomically-Guided Patch Extraction:

   • Instead of analyzing the whole image, the system extracts tissue patches from two clinically relevant regions defined by the segmentation masks:
     • Pancreas Patches: Non-overlapping patches sampled from within the pancreas mask, excluding the splenic vein.
     • Peri-venous Fat Patches: Patches sampled from the fat region immediately beneath the splenic vein contour.
   • This strategy automates the clinical comparison between pancreatic tissue and surrounding retroperitoneal fat.

3. Feature Engineering and Classification:

   • Feature Extraction: For each patch, a feature vector is computed comprising intensity statistics (mean, std, median), normalized intensity histograms, Laplacian variance (texture sharpness), local contrast, and Sobel gradient magnitude.
   • Patient-Level Aggregation: Rather than classifying individual patches, the framework computes pairwise L2 distances between all fat patches and all pancreas patches for a given patient. The patient-level feature vector aggregates statistics from this distance matrix (e.g., mean, percentiles, nearest-neighbor statistics) and the mean feature difference between the two tissue types.
   • Classification: The system evaluates both unsupervised (K-Means clustering) and supervised (SVM with RBF and linear kernels, Random Forest, Gradient Boosting, Logistic Regression, KNN) classifiers. The clinical intuition is that in a fatty pancreas, the echogenicity distributions of fat and pancreas overlap, resulting in smaller pairwise distances compared to a normal pancreas.

Key Contributions

• First End-to-End Pipeline: The authors present the first fully automated pipeline for fatty pancreas classification from ultrasound that integrates segmentation, anatomically-guided patch extraction, and texture-based classification.
• Domain-Specific Transfer Learning: The study demonstrates that initializing pancreas/splenic vein segmentation models with weights from a liver segmentation task substantially outperforms zero-shot and fine-tuned MedSAM (Segment Anything Model) on this specific task, particularly in a prompt-free setting.
• Clinical Mimicry: The feature engineering approach explicitly mimics clinical assessment by comparing the texture of peri-venous fat to pancreatic parenchyma, providing an interpretable signal for classification.
• Robust Validation: The full pipeline is validated using 5-fold cross-validation with independent retraining of segmentation models for each fold, ensuring results are not artifacts of a specific data split.

Experimental Results
The framework was validated on a clinical dataset of 214 abdominal ultrasound images, with 107 expert-labeled cases (53 fatty, 55 normal).

• Segmentation: The TransUNet models achieved mean Dice scores of 0.712 (±0.018) for the pancreas and 0.699 (±0.071) for the splenic vein. While fine-tuned MedSAM achieved slightly higher Dice scores (0.834/0.837), it required ground-truth bounding box prompts at inference, whereas the proposed TransUNet operates fully automatically. Zero-shot MedSAM performed poorly (Dice ~0.56).
• Classification:
  • Supervised: The SVM with an RBF kernel achieved the highest performance with a mean cross-validated accuracy of 89.7% ± 1.8% and an F1 score of 0.898 ± 0.019. The Cohen's kappa was 0.794, indicating substantial agreement with expert labels.
  • Unsupervised: The K-Means baseline achieved 87.8% ± 1.4% accuracy, demonstrating that the engineered features capture the relevant clinical signal even without labeled training data.
  • Robustness: The system showed notable robustness to segmentation errors; classification performance using predicted masks was comparable to, and in some cases slightly better than, using ground-truth masks. This suggests the pairwise texture comparison framework is tolerant to moderate boundary shifts.

Significance and Claims
The paper claims that this work represents the first end-to-end automated framework for fatty pancreas classification from ultrasound using segmentation-guided texture analysis. The authors emphasize that their approach successfully translates subjective clinical reasoning into a quantitative, automated workflow. They highlight that the proposed feature engineering effectively captures echogenicity differences, achieving stable performance despite a limited dataset size. The results suggest that further improvements in segmentation accuracy alone may not linearly translate to better classification, as the bottleneck may lie in dataset size or feature representation rather than segmentation precision. The authors acknowledge limitations regarding the single-center dataset, lack of inter-observer variability assessment, and the binary nature of the classification, noting that future work should focus on multi-center validation and extending the framework to multi-grade severity assessment.