Brain Tumor Type Classification via Capsule Networks (1802.10200v2)

Abstract: 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.

• The paper demonstrates that CapsNet achieves superior classification accuracy (86.56%) over traditional CNNs on brain tumor MRI data.
• It employs a CapsNet architecture that effectively handles rotation and affine transformations while mitigating overfitting with small datasets.
• Comparative experiments show that using segmented tumor regions enhances classification performance compared to whole brain scans.

Brain Tumor Type Classification via Capsule Networks

The paper addresses the challenging task of brain tumor type classification by using Capsule Networks (CapsNets), juxtaposing them against traditional Convolutional Neural Networks (CNNs). CapsNets are proposed as a viable alternative to CNNs owing to their robustness to rotation, affine transformations, and ability to work with smaller datasets—a significant advantage when considering medical imaging datasets like MRI, which often bear small sample sizes and varied orientations.

Methodology Evaluation

The paper sets aside conventional methods that rely heavily on feature extraction and pre-acquired knowledge. Instead, it employs a CapsNet architecture to determine brain tumor types, focusing not only on greater accuracy but also on resisting overfitting tendencies that arise from limited datasets. CNNs, while pervasive in image classification, struggle with transformations and necessitate large amounts of data—constraints that CapsNets potentially bypass. The research outlines several enhancements in architecture that yield varying performance outcomes:

• Adapting the number of feature maps and architectural layers for optimal performance.
• Comparing CapsNet's prediction accuracy when using either segmented tumor regions or whole brain images as input.

Experimental Insights

Experiments reveal CapsNet's superior performance against CNNs, showcasing a specifically tuned CapsNet architecture that achieves notable accuracy improvements. For instance, a CapsNet with a convolutional layer of 64 feature maps provides an accuracy of 86.56%, superior to alternative configurations scaled or architecturally akin to original Capsule propositions.

Numerical and Structural Outcomes

Quantified outcomes include an increased prediction accuracy with specific architectural configurations and learning paradigms. Utilizing early stopping to counter over-fitting is emphasized—every epoch's total loss shows an initial rapid decrease, substantially due to CapsNet loss.

The experimentation delineates that segmented tumor data leads to more accurate classifications than whole brain MRI inputs. This is predicated on CapsNet's underlying mechanism, which aims to map entire input features, thus encountering degradation with inconsistent background data inherent in whole brain scans.

Interpretative Analysis and Implications

CapsNet's inherent structure of 'routing by agreement' and the employment of activity vectors signifies a departure from traditional CNN pooling layers. The capsule approach retains transformation invariance and captures contextual spatial hierarchies, presenting pronounced improvements for medical image classifications. Visualization of ClassCaps outputs provides interpretative depth, revealing potential latent feature relationships such as tumor size and orientation.

Future Directions

Prospects for research extend into:

• Optimization of CapsNet architectures specifically tailored for medical contexts.
• Exploration of deeper capsular layers for potentially superior classification performance.
• Application to a wider variety of medical imaging modalities and tumor types to explore the broader efficacy of capsular treatments.

This research posits that CapsNets offer ha promising alternative solution for intricate classification tasks, especially in medical imaging domains. The capacity to withstand small dataset limitations while efficiently handling varied transformations paves the way for more effective, automated diagnostic criteria in clinical practices.