Source-linked AI summary
DeepOrgan: Multi-level Deep Convolutional Networks for Automated Pancreas Segmentation
Holger R. Roth, Le Lu, Amal Farag, Hoo-Chang Shin, Jiamin Liu, Evrim Turkbey, Ronald M. Summers
TL;DR
Automatic pancreas segmentation is challenging because pancreatic anatomy varies substantially across patients. The paper uses a coarse-to-fine, bottom-up pipeline of multi-level ConvNets on patches and regions, achieving 71.8±10.7% testing DSC in 4-fold cross-validation.
Problem
Pancreas segmentation lacks the consistently high accuracy achieved for several other organs because pancreatic anatomy varies drastically between patients.
Method
A bottom-up coarse-to-fine pipeline progressively classifies image patches and superpixel regions with P-, R1-, and stacked R2-ConvNets, followed by probability smoothing and structured prediction.
Results
71.8±10.7% testing DSC was achieved on CT scans from 82 patients using 4-fold cross-validation.
Takeaways & Limitations
The approach reports the highest average testing DSC in the paper and is presented as potentially generalizable to segmentation problems with large variations and pathologies.
Takeaways & Limitations
Supervoxel-based regional ConvNets require at least one-order-of-magnitude wider input layers and significantly more parameters to train.
Abstract
from arXiv · showhide
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 previous segmentation methods from achieving high accuracies, especially compared to other organs such as the liver, heart or kidneys. In this paper, we present a probabilistic bottom-up approach for pancreas segmentation in abdominal computed tomography (CT) scans, using multi-level deep convolutional networks (ConvNets). We propose and evaluate several variations of deep ConvNets in the context of hierarchical, coarse-to-fine classification on image patches and regions, i.e. superpixels. We first present a dense labeling of local image patches via $P{-}\mathrm{ConvNet}$ and nearest neighbor fusion. Then we describe a regional ConvNet ($R_1{-}\mathrm{ConvNet}$) that samples a set of bounding boxes around each image superpixel at different scales of contexts in a "zoom-out" fashion. Our ConvNets learn to assign class probabilities for each superpixel region of being pancreas. Last, we study a stacked $R_2{-}\mathrm{ConvNet}$ leveraging the joint space of CT intensities and the $P{-}\mathrm{ConvNet}$ dense probability maps. Both 3D Gaussian smoothing and 2D conditional random fields are exploited as structured predictions for post-processing. We evaluate on CT images of 82 patients in 4-fold cross-validation. We achieve a Dice Similarity Coefficient of 83.6$\pm$6.3% in training and 71.8$\pm$10.7% in testing.
1 Introduction
Pancreas segmentation remains difficult because the organ varies drastically in shape, size, and location, despite strong results for several other organs.
- Pancreas segmentation supports quantitative organ volume analysis in CADx systems and may support pancreatic-cancer detection in CADe methods.
- Other abdominal organs commonly achieve DSCs above 90%, whereas automatic pancreas segmentation remains challenging.
- The pancreas exhibits drastic interpatient variation in shape, size, and abdominal location.
2 Methods
The method uses a hierarchical, coarse-to-fine pipeline that progresses from dense patch labeling to multi-scale superpixel classification for pancreas segmentation. It combines CT intensities, patch-level probability maps, multi-scale context, and structured post-processing.
- Coarse-to-fine segmentation: The pipeline performs coarse-to-fine pancreas segmentation by progressively labeling image patches, regions, and the entire organ from an initial random-forest superpixel set.It uses a bottom-up strategy rather than top-down multi-atlas registration and label fusion.
- Regional classification: R1-ConvNet classifies each candidate superpixel using CT-intensity bounding boxes sampled at multiple spatial scales in a zoom-out sequence.Increasing context helps address the complex anatomy surrounding pancreas candidates.
- Patch labeling: P-ConvNet densely labels 2.5D patches extracted across initial superpixel regions, then uses nearest-neighbor interpolation to form the patch-level probability map P0.Patches combine axial, coronal, and sagittal planes and are sampled every n voxels for efficiency.
- Regional classification: R2-ConvNet stacks CT intensities with the P0 patch-response map as an additional channel, while irregular superpixels are warped into regular square inputs.The two regional variants differ in whether they use intensity alone or intensity plus patch-level responses.
- Structured prediction: At testing, multi-scale regional probabilities are averaged per superpixel, assigned to pixels or voxels, and refined with 3D Gaussian filtering and optional 2D CRF boundary modeling.The CRF adds a neighboring-superpixel edge term to the regional R2-ConvNet term for min-cut/max-flow segmentation.
3 Results & Discussion
Using 4-fold cross-validation, the method combines augmented training, Gaussian-smoothed ConvNet probability maps, and structured prediction to evaluate pancreas segmentation. Testing performance reaches 71.8% average DSC, with smoothing substantially improving results over candidate-region generation.
- 82 CT volumes were evaluated with 4-fold cross-validation, using random hard splits with 21, 21, 20, and 20 patients in the testing folds.
- Training uses scale and random TPS deformation augmentation with Ns = 4 and Nt = 8, while testing uses Ns = 4 and a 3-voxel Gaussian kernel.
- 71.8% testing average DSC was achieved with stacked R2-ConvNet and Gaussian smoothing, improving 45.7% over candidate-region generation at 26.1%.
- Gaussian 3D smoothing markedly improved average DSCs, with calibrated operation points of p0 = 0.2, p1 = 0.5, and p2 = 0.6 in training.
- G(P0(x)) achieved 69.5% mean DSC and was more efficient, while G(P2(x)) improved surface-to-surface distance from 1.46±1.5mm to 0.94±0.6mm.
- The method achieved 71.8±10.7% testing DSC versus 83.6±6.3% training DSC, while CRF(P2(x)) reached 68.2% ±4.1%.
4 Conclusion
The paper presents a coarse-to-fine, bottom-up pancreas segmentation method using multi-level deep ConvNets. It reports 71.8±10.7% testing DSC at a computational cost of a few minutes and discusses broader applicability.
- The proposed approach performs pancreas segmentation in abdominal CT scans using a bottom-up, coarse-to-fine hierarchy of multi-level deep ConvNets.
- 71.8±10.7% testing DSC and 83.6±6.3% training DSC are reported at a computational cost of a few minutes rather than hours.
- The approach can be incorporated into multi-organ segmentation frameworks by specifying additional tissue types because ConvNets support multi-class classification.
- The authors state that the deep-learning organ-segmentation approach could generalize to other segmentation problems involving large variations and pathologies, including tumors.