Source-linked AI summary
Deep convolutional networks for pancreas segmentation in CT imaging
Holger R. Roth, Amal Farag, Le Lu, Evrim B. Turkbey, Ronald M. Summers
TL;DR
Automatic pancreas segmentation is difficult because anatomical variability and weak CT contrast complicate delineation, despite its role in computer-aided diagnosis. The paper uses hierarchical superpixel classification with random forests and ConvNets, achieving a 68% ± 10% average testing Dice score on 20 CT cases and reporting performance comparable to recent methods.
Problem
Pancreas segmentation matters for computer-aided diagnosis, but anatomical variability and low boundary contrast make accurate automatic segmentation challenging.
Method
The method hierarchically classifies SLIC superpixels, first retaining high-probability regions with random forests, then refining them with a multiscale ConvNet and 3D Gaussian smoothing.
Results
68% ± 10% average Dice score (range, 43-80%) was achieved on 20 testing CT cases.
Takeaways & Limitations
ConvNets can be generalized to pancreas segmentation by classifying superpixels into pancreas and non-pancreas tissue types.
Abstract
from arXiv · showhide
Automatic organ segmentation is an important prerequisite for many computer-aided diagnosis systems. The high anatomical variability of organs in the abdomen, such as the pancreas, prevents many segmentation methods from achieving high accuracies when compared to other segmentation of organs like the liver, heart or kidneys. Recently, the availability of large annotated training sets and the accessibility of affordable parallel computing resources via GPUs have made it feasible for "deep learning" methods such as convolutional networks (ConvNets) to succeed in image classification tasks. These methods have the advantage that used classification features are trained directly from the imaging data. We present a fully-automated bottom-up method for pancreas segmentation in computed tomography (CT) images of the abdomen. The method is based on hierarchical coarse-to-fine classification of local image regions (superpixels). Superpixels are extracted from the abdominal region using Simple Linear Iterative Clustering (SLIC). An initial probability response map is generated, using patch-level confidences and a two-level cascade of random forest classifiers, from which superpixel regions with probabilities larger 0.5 are retained. These retained superpixels serve as a highly sensitive initial input of the pancreas and its surroundings to a ConvNet that samples a bounding box around each superpixel at different scales (and random non-rigid deformations at training time) in order to assign a more distinct probability of each superpixel region being pancreas or not. We evaluate our method on CT images of 82 patients (60 for training, 2 for validation, and 20 for testing). Using ConvNets we achieve average Dice scores of 68%+-10% (range, 43-80%) in testing. This shows promise for accurate pancreas segmentation, using a deep learning approach and compares favorably to state-of-the-art methods.
1. INTRODUCTION
Pancreas segmentation supports computer-aided diagnosis, but substantial anatomical and imaging variability makes accurate automatic segmentation difficult. The paper aims to reproduce expert manual segmentations using computer vision and medical image computing.
- Pancreas segmentation is an important input for computer-aided diagnosis systems providing quantitative analysis.
- Accurate pancreas segmentation is challenging because its shape, size, and location vary drastically between patients.
- Variable visceral fat contrast and low boundary contrast in CT further complicate pancreas delineation.
- The study aims to replicate experienced-radiologist pancreas segmentations using computer vision and medical image computing.
2. METHODS
The method uses hierarchical coarse-to-fine classification of abdominal superpixels, combining random forests with a ConvNet to estimate pancreas probabilities. Multi-scale sampling, deformation-based augmentation, and 3D smoothing refine the final segmentation.
- Data-driven learning: The approach trains ConvNet features directly from imaging data rather than relying on hand-crafted features designed from human experience.The paper situates this approach within the increased availability of annotated datasets and affordable GPU computing.
- Hierarchical candidate generation: Abdominal superpixels are extracted with SLIC, then patch-level features and a two-level random-forest cascade generate a probability response map.High-probability regions are retained as candidate pancreas superpixels.
- ConvNet classification: Retained superpixels are classified by a ConvNet using bounding boxes sampled at multiple scales and random non-rigid training deformations.The deformations are generated with thin-plate splines to model plausible physical variation and help avoid overfitting.
- ConvNet architecture: The ConvNet uses five convolutional layers with max-pooling, locally fully connected layers, dropout connections, and a final two-way softmax for pancreas versus non-pancreas.Learned filters include shape-enhancement and texture filters.
- Probability refinement: ConvNet probabilities are assigned to pixels within each superpixel and then smoothed with a Gaussian filter across neighboring regions and CT slices.The 3D filtering produces a smoother probability map and propagates 2D probabilities through local 3D neighborhoods.
3. RESULTS
Testing used 20 CT cases after training on 60 and validating on 2, with the proposed ConvNet improving pancreas segmentation over initial superpixel labeling. Three-dimensional smoothing and multiple scales were evaluated, while testing required only 1–3 minutes per CT volume.
- Reference performance: 81% ± 3% average Dice (range, 75–89%) was the optimally achievable superpixel classification on the testing set.This upper bound reflects the ability of superpixels to capture pancreas boundaries.
- Testing performance: 68% ± 10% Dice (range, 43–80%) was achieved on the 20-case test set using the selected ConvNet operating point.The operating point used pConvNet = 0.4 after 3D Gaussian smoothing with Ns = 4.
- Testing performance: 41% improvement in Dice was obtained over the initial superpixel candidate labeling, which achieved 27% ± 6% Dice (range, 16–42%).The initial labeling maintained high pancreas sensitivity through over-segmentation.
- Scale and smoothing effects: 56% to 68% mean Dice improvement resulted from applying proposed 3D smoothing to ConvNet probabilities from 2D slices.Figure 6 compares unsmoothed and 3D-smoothed probabilities at Ns = 1 and Ns = 4.
- Computational cost: 1 to 3 minutes per CT volume was required for testing, depending on the number of scales Ns.Training used 855,500 64 × 64-pixel superpixel images and took 55 hours for 100 epochs on a GPU.
4. CONCLUSIONS
The study demonstrates ConvNet-based superpixel classification for automated pancreas segmentation and reports promising results despite a relatively simple approach. The method could also be extended to multi-organ segmentation through multi-class tissue classification.
- Conclusions: ConvNets can be generalized to medical image analysis tasks such as pancreas segmentation.The proposed system classifies superpixels into pancreas and non-pancreas tissue types.
- Conclusions: 68.8% was the upper end of reported average Dice scores from recent pancreas-segmentation methods, while this study used a random 60/2/20 training, validation, and testing split.The paper reports average Dice scores ranging from 46.6% to 68.8% for recent methods and describes its own results as similar or better.
- Future application: Multi-organ segmentation is a supported extension because ConvNets can classify more tissue types through multi-class classification.The paper proposes specifying additional tissue types using a similar approach.