Source-linked AI summary
Brain Tumor Type Classification via Capsule Networks
Parnian Afshar, Arash Mohammadi, Konstantinos N. Plataniotis
TL;DR
Brain tumor type classification requires accurate early identification despite CNN limitations with transformations and small datasets. The paper designs and evaluates CapsNets across architectures and input types, finding that they outperform CNNs and perform better on segmented tumors than whole-brain images.
Problem
Brain tumor type classification is clinically important, while CNNs have limitations with input transformations and small medical-image datasets.
Method
The paper designs a CapsNet architecture, addresses over-fitting with early stopping, compares whole-brain and segmented-tumor inputs, and visualizes learned features through reconstruction.
Results
CapsNets outperform CNNs for brain tumor classification and perform better with segmented tumors than whole-brain images.
Takeaways & Limitations
CapsNets provide an effective approach for brain tumor classification with small training samples and can improve accuracy through convolutional feature-map design.
Takeaways & Limitations
CapsNets can over-fit relatively small datasets, requiring careful training and early stopping for reasonable generalization.
Abstract
from arXiv · showhide
Brain tumor is considered as one of the deadliest and most common form of cancer both in children and in adults. Consequently, determining the correct type of brain tumor in early stages is of significant importance to devise a precise treatment plan and predict patient's response to the adopted treatment. In this regard, there has been a recent surge of interest in designing Convolutional Neural Networks (CNNs) for the problem of brain tumor type classification. However, CNNs typically require large amount of training data and can not properly handle input transformations. Capsule networks (referred to as CapsNets) are brand new machine learning architectures proposed very recently to overcome these shortcomings of CNNs, and posed to revolutionize deep learning solutions. Of particular interest to this work is that Capsule networks are robust to rotation and affine transformation, and require far less training data, which is the case for processing medical image datasets including brain Magnetic Resonance Imaging (MRI) images. In this paper, we focus to achieve the following four objectives: (i) Adopt and incorporate CapsNets for the problem of brain tumor classification to design an improved architecture which maximizes the accuracy of the classification problem at hand; (ii) Investigate the over-fitting problem of CapsNets based on a real set of MRI images; (iii) Explore whether or not CapsNets are capable of providing better fit for the whole brain images or just the segmented tumor, and; (iv) Develop a visualization paradigm for the output of the CapsNet to better explain the learned features. Our results show that the proposed approach can successfully overcome CNNs for the brain tumor classification problem.
1. INTRODUCTION
Brain tumor type classification is clinically important but difficult, motivating automated MRI-based approaches. CapsNets are adopted to address CNN limitations, compare whole-brain and segmented-tumor inputs, and improve feature explainability.
- Early tumor-type identification supports more precise treatment planning and prediction of patient response.
- Human inspection is time-consuming and error-prone because it depends heavily on radiologist experience and skill.
- MRI is favored for brain tumor classification because of its harmless nature.
- CNNs require varied transformation examples and often perform poorly on small medical-image datasets.
- CapsNets use pose-aware activity vectors and routing by agreement, activating higher-level Capsules when predictions agree.
- The study explores CapsNet architectures, whole-brain versus segmented-tumor inputs, over-fitting, and visualization of learned features.
2. PROBLEM FORMULATION
The task is to classify MRI brain tumors into three categories using Capsules instead of conventional neurons. The section reviews CNN design properties and motivates CapsNets through CNN limitations involving location information and transformations.
- 2. PROBLEM FORMULATION: The goal is to classify test MRI images into Meningioma, Pituitary, or Glioma using a deep learning architecture.
- 2. PROBLEM FORMULATION: The proposed architecture uses Capsules instead of neurons for brain tumor type classification.
- 2.1. Convolutional Neural Networks: CNN layers process local neighborhoods to extract elementary features such as edges and corners.
- 2.1. Convolutional Neural Networks: Later CNN layers combine elementary features to detect higher-order features.
- 2.1. Convolutional Neural Networks: CNNs share similar feature detectors across the entire image through shared weights.
- 2.1. Convolutional Neural Networks: CNN sub-sampling layers treat precise feature locations as potentially harmful because locations vary across instances.
- 2.1. Convolutional Neural Networks: Sub-sampling provides limited translational invariance while losing the exact locations of the most active feature detectors.
- 2.1. Convolutional Neural Networks: Capsule networks are introduced as more robust to translation and rotation than CNNs.
3. CAPSNETS FOR BRAIN TUMOR CLASSIFICATION
The paper designs a Capsule Network for brain tumor classification, replacing conventional neural units and pooling with capsules and routing by agreement. The proposed architecture uses a reduced convolutional feature-map count, MRI inputs, class capsules, reconstruction, and early stopping to address over-fitting.
- Capsule network properties: Capsules encode pose parameters in activity vectors, whose lengths represent the probability that corresponding entities exist.Capsule outputs therefore retain information about position, orientation, scaling, and skewness.
- Capsule network properties: Routing by agreement increases coupling between capsules when lower-level predictions conform to higher-level capsule outputs.Each capsule predicts parent-capsule outputs, and agreement determines the coupling coefficients.
- Designed CapsNet: The network processes down-sampled MRI images through convolutional and primary-capsule layers before three 16-dimensional class capsules represent the tumor categories.The class capsules correspond to Meningioma, Pituitary, and Glioma.
- Designed CapsNet: A decoder with 512, 1024, and 4096 neurons reconstructs the input image by minimizing squared differences between the input and reconstruction.The final decoder layer has as many neurons as the pixels in the input image.
- Training and input choice: CapsNet training over-fits on the relatively small MRI dataset, so the authors use validation-based early stopping to improve generalization.Training is stopped when validation accuracy begins to decrease.
4. EXPERIMENTAL SETUP
Experiments evaluated Capsule-network architectures, training losses, CNN comparisons on whole-brain and segmented-tumor MRI inputs, and feature visualization. The proposed CapsNet outperformed CNNs for both input types, while segmented tumors yielded better CapsNet performance than whole-brain images.
- Architecture selection: Reducing convolutional feature maps from 256 to 64 produced the highest accuracy among the tested Capsule-network architectures.
- Training: The proposed architecture was trained for 10 epochs using CapsNet and decoder losses, with total loss mostly dependent on the CapsNet loss.
- Classification comparison: CapsNet outperformed CNN for both whole-brain and segmented-tumor images, although whole-brain images produced lower CapsNet accuracy than segmented tumors.
- Feature visualization: Tweaking ClassCaps activation vectors enabled reconstruction-based visualization of learned features, including tumor size and width.
5. CONCLUSION
The study applied Capsule networks to brain tumor type classification and found advantages over CNNs, including stronger performance on segmented tumors and benefits from architectural changes. Future work will examine whether additional capsule layers improve accuracy.
- Capsule networks outperformed CNNs for brain tumor classification while handling small training sets and using equivariant units.
- Changing the convolutional feature-map count increased the Capsule network’s classification accuracy.
- Capsule networks performed better on segmented tumors than on whole-brain images in the experiments.
- Future work will investigate the effect of adding more capsule layers on classification accuracy.