Source-linked AI summary
Fast CapsNet for Lung Cancer Screening
Aryan Mobiny, Hien Van Nguyen
TL;DR
Detecting lung nodules in CT is a major screening challenge, while CNNs are state of the art and CapsNet performance on medical images remains unclear. This paper evaluates CapsNets for 3D lung nodule classification and proposes consistent routing and a convolutional decoder. CapsNets outperform CNNs on small datasets, and modified routing is reported as 3× faster while maintaining accuracy.
Problem
Lung nodule detection is a major screening challenge, and CapsNet performance on medical imaging tasks remains unclear while dynamic routing is computationally expensive for high-dimensional CT scans.
Source: §1 · pp. 1–2Method
The paper evaluates CapsNets for 3D lung nodule classification and proposes consistent routing plus a convolutional decoder that reconstructs from PrimaryCaps predictions and class information.
Source: §1 · p. 2, §3 · pp. 4–5Results
Takeaways & Limitations
Abstract
from arXiv · showhide
Lung cancer is the leading cause of cancer-related deaths in the past several years. A major challenge in lung cancer screening is the detection of lung nodules from computed tomography (CT) scans. State-of-the-art approaches in automated lung nodule classification use deep convolutional neural networks (CNNs). However, these networks require a large number of training samples to generalize well. This paper investigates the use of capsule networks (CapsNets) as an alternative to CNNs. We show that CapsNets significantly outperforms CNNs when the number of training samples is small. To increase the computational efficiency, our paper proposes a consistent dynamic routing mechanism that results in $3\times$ speedup of CapsNet. Finally, we show that the original image reconstruction method of CapNets performs poorly on lung nodule data. We propose an efficient alternative, called convolutional decoder, that yields lower reconstruction error and higher classification accuracy.
1 Introduction
Lung nodule detection is a major challenge in screening, and although CNNs are state of the art for classification, CapsNet performance on medical imaging and the cost of routing remain open concerns. The paper investigates CapsNets for lung nodule classification and proposes changes to routing and reconstruction.
1 Introduction
- CNNs are state-of-the-art for lung nodule classification, but pooling discards location and pose information that may be useful for classification.The paper identifies routing as the source of this drawback.Source: §1 · p. 1
- CapsNets encode relationships such as the locations, scales, and orientations of parts relative to the whole object.This representation is described as supporting recognition from 3D views not seen during training.Source: §1 · p. 1
- CapsNet performance on medical image analysis and the comparison with state-of-the-art methods remain unclear, while costly routing limits use on high-dimensional CT scans.The paper specifically identifies 3D CT scans as a high-dimensional setting.Source: §1 · p. 2
- CapsNets significantly outperform CNNs on small datasets, while the proposed routing is 3× faster and the convolutional decoder improves reconstruction and classification.The paper also reports that CapsNets compare favorably to CNNs with larger training sets.Source: §1 · p. 2
2 Capsule Network
Capsules represent object presence and pose in activation vectors, and dynamic routing sends predictions to higher-level capsules that agree with them. The network combines this encoder with a reconstruction decoder.
2 Capsule Network
- A capsule is a group of neurons whose activation vector encodes object presence through its length and pose information through its direction.The length represents the estimated probability that the object exists; direction captures properties such as size, orientation, and position.Source: §2 · p. 2
- A learned linear transformation maps each lower-layer capsule to a prediction about a higher-layer capsule, encoding their part–whole relationship.Dynamic routing iteratively adjusts routing coefficients so child capsules send information to parent capsules that agree with their predictions.Source: §2 · p. 3
- Each parent capsule forms a weighted sum of child predictions and applies a squash non-linearity that limits output length to 1 without changing direction.The output-vector length can therefore be interpreted as the probability that a feature is detected.Source: §2 · p. 3
- The encoder uses convolutional and capsule layers, with one 16-dimensional DigitCaps capsule per class whose vector lengths determine the predicted class.Routing occurs between the PrimaryCaps and DigitCaps layers.Source: §2 · pp. 3–4
- The decoder reconstructs the input from the final capsules, encouraging information preservation across the network and acting as a regularizer against over-fitting.It masks out all final-capsule outputs except those for the target class during training or the predicted class during testing.Source: §2 · p. 4
3 Fast Capsule Network
FastCapsNet reduces routing operations by sharing routing coefficients among capsules at the same pixel location, and replaces the original reconstruction decoder with a convolutional alternative. The proposed routing is reported as faster while maintaining similar classification accuracy.
3 Fast Capsule Network
- The proposed architecture is 3× faster than CapsNet while achieving similar classification accuracy.Capsules at the same pixel location share routing coefficients; the method uses one 256D capsule per location instead of 32 8D capsules.Source: §3 · p. 4
- Sharing routing coefficients among capsules at the same pixel location reduces their number by a factor of 32 in the original 32-capsule-per-location arrangement.The constraint uses one capsule at each pixel location and increases capsule dimensionality from 8D to 256D.Source: §3 · p. 4
- The original feed-forward decoder reconstructs lung nodule data poorly, which the authors suggest may reflect high visual variability within the data.The convolutional decoder instead uses target-class PrimaryCaps prediction vectors and a replicated one-hot class vector as input, followed by two transposed-convolution layers; it has about 5K parameters.Source: §3 · pp. 4–5
4 Experiments
Experiments compare capsule networks with CNNs on lung-nodule classification, including different training-set sizes and 2D/3D inputs. FastCapsNet performs better in the reported comparisons and reduces routing costs.
4 Experiments
- The study used 226 CT scans, with candidate nodules cropped and resized to 32×32×32 pixels and reviewed by at least one board-certified radiologist.Scans came from General Electric and Siemens scanners.Source: §4 · p. 5
- For 2D inputs, FastCapsNet uses 64 PrimaryCaps instead of the original CapsNet’s 2,048, reducing voting coefficients by 32 times.The 2D inputs were middle slices along the x-axis; the 3D version uses 512 PrimaryCaps rather than 23,328.Source: Figure 3 · p. 6, §4 · p. 7
- FastCapsNet has fewer parameters than the other models and runs more than 3 times faster than the original CapsNet in training and testing.Capsule networks also achieve relatively lower error rates than AlexNet and ResNet.Source: Table 1 · p. 7, §4 · p. 8
- The original 3D CapsNet failed to converge, whereas 3D FastCapsNet reached less than 20% test loss after 3 training epochs.The 3D architecture uses 512 PrimaryCaps, compared with 23,328 in the original network.Source: §4 · pp. 7–8
- The 3D FastCapsNet beats the 2D version by almost 2% accuracy, while capsule networks stay about 70% accurate with very few samples as AlexNet and ResNet-50 approach chance.The precision-recall curves, sample-size accuracy, and train/test loss are shown in Fig. 4.Source: Figure 4 · p. 7, §4 · p. 8
5 Conclusions
The authors conclude that CapsNet is a promising alternative to CNNs for lung-nodule classification, especially with small training sets.
5 Conclusions
- CapsNets compare favorably to CNNs with large training sets and significantly outperform them on small datasets.The authors identify unsupervised learning with CapsNets as future work.Source: §5 · p. 8
- Modifying the routing mechanism speeds up CapsNets 3 times while maintaining the same accuracy.This is the authors’ reported conclusion about the modified routing mechanism.Source: §5 · p. 8
- The authors describe CapsNet as a promising alternative to CNNs for lung-nodule classification.Source: §5 · p. 8