Source-linked AI summary
Spatial Aggregation of Holistically-Nested Networks for Automated Pancreas Segmentation
Holger R. Roth, Le Lu, Amal Farag, Andrew Sohn, Ronald M. Summers
TL;DR
Pancreas segmentation is difficult because anatomical variability limits accuracy. The paper combines holistically nested CNN interior and boundary cues with random-forest spatial aggregation, achieving 78.01% ± 8.2% testing DSC on 82 CT patients in 4-fold cross-validation.
Problem
Pancreas anatomical variability limits automatic segmentation accuracy relative to several other abdominal organs.
Method
The method combines deeply learned pancreas interior and boundary cues with weighted fusion and random-forest spatial aggregation.
Results
78.01% ± 8.2% testing DSC was achieved on CT scans of 82 patients using 4-fold cross-validation.
Takeaways & Limitations
Spatially aggregating learned interior, boundary, and location information advances the reported pancreas-segmentation performance under the evaluated protocol.
Takeaways & Limitations
Candidate-region generation is treated as fixed and is not the focus of the study, although other bounding-box detection methods could replace it.
Abstract
from arXiv · showhide
Accurate automatic organ segmentation is an important yet challenging problem for medical image analysis. The pancreas is an abdominal organ with very high anatomical variability. This inhibits traditional segmentation methods from achieving high accuracies, especially compared to other organs such as the liver, heart or kidneys. In this paper, we present a holistic learning approach that integrates semantic mid-level cues of deeply-learned organ interior and boundary maps via robust spatial aggregation using random forest. Our method generates boundary preserving pixel-wise class labels for pancreas segmentation. Quantitative evaluation is performed on CT scans of 82 patients in 4-fold cross-validation. We achieve a (mean $\pm$ std. dev.) Dice Similarity Coefficient of 78.01% $\pm$ 8.2% in testing which significantly outperforms the previous state-of-the-art approach of 71.8% $\pm$ 10.7% under the same evaluation criterion.
1 Introduction
Pancreas segmentation remains difficult because the organ varies substantially in shape, size, and location, limiting reported accuracy below that achieved for several other abdominal organs. CNN-based end-to-end representation learning offers a promising basis for pixel-level semantic segmentation.
- Pancreas variability in shape, size, and location limits reported segmentation accuracy to below 73% DSC, whereas liver, heart, and kidney segmentation exceeds 90% DSC.
- The central challenge is assigning accurate pixel-level organ labels despite the pancreas’s high anatomical variability.
- Deep CNNs learn task-relevant feature representations end-to-end and have improved pixel-level semantic segmentation in medical imaging.
2 Methods
The method combines holistically nested CNN predictions of pancreas interiors and boundaries with multiscale object proposals and random-forest spatial aggregation. It trains these cues on whole CT images and uses learned fusion, class-balanced losses, and superpixel features to produce segmentation labels.
- The method integrates deeply learned organ-interior and boundary cues with subsequent spatial aggregation for holistic CT organ segmentation.
- 2.2.1 Holistically-Nested Nets:: Multiple HNN side outputs capture different receptive-field scales and are combined through a learned weighted-fusion layer.
- 2.2 Semantic Mid-level Segmentation Cues: HNN independently produces pancreas interior and boundary prediction maps from whole images using binary pixel-labeling supervision.
- 2.2.2 Network formulation:: Training uses side-output losses with per-pixel class balancing to offset the imbalance between positive organ or edge pixels and negative pixels.
- 2.3 Learning Organ-specific Segmentation Object Proposals: MCG applies multiscale grouping to HNN-B boundary maps to generate organ-specific superpixels and merged object proposals.
- Random-forest aggregation represents each superpixel with CT, interior, boundary, and normalized spatial features.
3 Results & Discussion
Across 4-fold cross-validation, HNN-RF spatial aggregation achieved the strongest reported pancreas-segmentation performance, while comparisons with other methods depend on evaluation protocol and dataset.
- The evaluation used 82 contrast-enhanced CT volumes with random splits of approximately 60 training and 20 unseen testing patients in 4-fold cross-validation.This protocol differs from the leave-one-patient-out evaluation commonly used in earlier work.
- 78.01% average DSC was achieved by HNN-RF, significantly improving over HNN-I’s 76.99% under 4-fold cross-validation.The improvement was statistically significant with p<0.001.
- DCRF produced 77.14% average DSC and did not noticeably improve upon HNN-I, unlike the proposed spatial aggregation.
- HNN-RF reported 78.01% average DSC with 8.2% standard deviation, compared with previous reported standard deviations of 18.6% and 15.3%.The minimal HNN-RF DSC was 34.11%, whereas several earlier methods reported cases below 10%.
- Direct comparison with most previous methods is limited because they used different datasets, although the proposed result was the highest reported under the same 4-fold criterion.
4 Conclusion
The paper concludes that combining learned organ-interior and boundary cues through spatial aggregation improves pancreas segmentation. The approach achieved 78.01%±8.2% DSC in testing at a computational cost of 2–3 minutes.
- 78.01%±8.2% DSC in testing was achieved by combining learned interior and boundary cues through spatial aggregation.
- The proposed approach required 2–3 minutes of computation rather than the hours reported for several earlier methods.
- The authors suggest that the deep-learning organ-segmentation approach could generalize to problems with substantial anatomical variation and pathologies, including tumors.