Source-linked AI summary
COVID-CAPS: A Capsule Network-based Framework for Identification of COVID-19 cases from X-ray Images
Parnian Afshar, Shahin Heidarian, Farnoosh Naderkhani, Anastasia Oikonomou, Konstantinos N. Plataniotis, Arash Mohammadi
TL;DR
COVID-19 diagnosis requires rapid, accurate methods, while existing testing, human interpretation, and CNN-based approaches face sensitivity, specificity, spatial-information, and data limitations. The paper introduces COVID-CAPS, a Capsule Network framework for X-ray classification, achieving strong performance and further gains from external X-ray pre-training.
Problem
Rapid COVID-19 diagnosis is important, but RT-PCR has limited sensitivity and accessibility, imaging findings overlap with other conditions, and CNNs can lose spatial information while requiring large datasets.
Method
COVID-CAPS uses convolutional and Capsule layers with routing by agreement to identify COVID-19 from X-ray images, including pre-training on an external X-ray dataset.
Results
COVID-CAPS achieved Accuracy 95.7%, Sensitivity 90%, Specificity 95.8%, and AUC 0.97; external X-ray pre-training increased accuracy to 98.3% and specificity to 98.6%, with sensitivity of 80%.
Takeaways & Limitations
The reported results support COVID-CAPS as a low-parameter X-ray-based framework with further performance gains from similar-domain pre-training.
Abstract
from arXiv · showhide
Novel Coronavirus disease (COVID-19) has abruptly and undoubtedly changed the world as we know it at the end of the 2nd decade of the 21st century. COVID-19 is extremely contagious and quickly spreading globally making its early diagnosis of paramount importance. Early diagnosis of COVID-19 enables health care professionals and government authorities to break the chain of transition and flatten the epidemic curve. The common type of COVID-19 diagnosis test, however, requires specific equipment and has relatively low sensitivity. Computed tomography (CT) scans and X-ray images, on the other hand, reveal specific manifestations associated with this disease. Overlap with other lung infections makes human-centered diagnosis of COVID-19 challenging. Consequently, there has been an urgent surge of interest to develop Deep Neural Network (DNN)-based diagnosis solutions, mainly based on Convolutional Neural Networks (CNNs), to facilitate identification of positive COVID-19 cases. CNNs, however, are prone to lose spatial information between image instances and require large datasets. The paper presents an alternative modeling framework based on Capsule Networks, referred to as the COVID-CAPS, being capable of handling small datasets, which is of significant importance due to sudden and rapid emergence of COVID-19. Our results based on a dataset of X-ray images show that COVID-CAPS has advantage over previous CNN-based models. COVID-CAPS achieved an Accuracy of 95.7%, Sensitivity of 90%, Specificity of 95.8%, and Area Under the Curve (AUC) of 0.97, while having far less number of trainable parameters in comparison to its counterparts. To further improve diagnosis capabilities of the COVID-CAPS, pre-training based on a new dataset constructed from an external dataset of X-ray images. Pre-training with a dataset of similar nature further improved accuracy to 98.3% and specificity to 98.6%.
1. Introduction
The introduction motivates rapid, accurate COVID-19 diagnosis from imaging and identifies limitations of existing tests, human interpretation, CNNs, and small medical datasets. It presents COVID-CAPS, a Capsule Network framework using X-ray images, and reports strong classification results with external-image pre-training.
- Early COVID-19 identification supports rapid patient isolation, potentially breaking transmission and flattening the epidemic curve.
- RT-PCR has relatively low sensitivity, requires specialized equipment, and is time-consuming, limiting rapid diagnosis.
- COVID-19 imaging findings can overlap with other infections and conditions, reducing specificity and complicating human-centered diagnosis.
- CNN-based approaches can lose spatial relations and therefore require large datasets containing diverse transformations, which are scarce for newly emerged COVID-19.
- COVID-CAPS applies Capsule Networks to COVID-19 identification from X-ray images and reports Accuracy 95.7%, Sensitivity 90%, Specificity 95.8%, and AUC 0.97.
- 98.3% accuracy and 98.6% specificity were achieved after pre-training on an external X-ray dataset, while sensitivity was 80%.
2. Capsule Networks
Capsule Networks represent image instances through capsules and use routing by agreement to combine predictions into capsule outputs. Their loss function separates present and absent classes using capsule lengths and class-dependent margins.
- Each capsule represents an image instance at a location, and its length determines the associated instance’s existence probability.
- Capsule i predicts the next layer’s capsules using a trainable weight matrix, producing prediction vectors for downstream capsules.
- Routing by agreement weights predictions according to their agreement with outputs, determining each downstream capsule’s contribution.
- Routing by agreement distinguishes CapsNets from CNNs by helping identify spatial relations.
- The CapsNet loss combines present-class and absent-class penalties using capsule lengths, margins m+ and m−, and hyperparameter λ.
3. The Proposed COVID-CAPS
COVID-CAPS combines convolutional and Capsule layers for X-ray-based COVID-19 identification, using routing by agreement and a class-imbalance-aware loss. The framework also considers pre-training on similar X-ray images rather than natural images.
- COVID-CAPS contains 4 convolutional layers and 3 Capsule layers, with the final Capsule layer representing positive and negative COVID-19 classes.The network receives 3D X-ray images and uses routing by agreement across its Capsule layers.
- The lengths of the final two Capsules represent the probabilities that each COVID-19 class is present.
- Because positive cases are fewer than negative cases, the loss function gives positive samples more weight based on class proportions.The paper defines loss+ for positive samples and loss− for negative samples.
- Pre-training uses an external X-ray dataset because its image nature is more similar to COVID-19 X-rays than ImageNet's natural images.
- Training uses Adam with an initial learning rate of 10^-3, 100 epochs, and batch size 16, followed by validation-based model selection and testing.
4. Experimental Results
Experiments binarize four X-ray labels into COVID-19-positive and negative classes and evaluate COVID-CAPS on classification performance, false positives, comparison models, and external-data pre-training.
- The dataset contains Normal, Bacterial, Non-COVID Viral, and COVID-19 labels, with the first three combined into the negative class.
- 95.7% accuracy, 90% sensitivity, 95.8% specificity, and 0.97 AUC were achieved by COVID-CAPS on the binarized X-ray dataset.
- 54% of false positives were normal cases, compared with 27% bacterial and 19% non-COVID cases.
- COVID-CAPS outperformed Reference in accuracy and specificity while using 295,488 trainable parameters versus 23 million.Reference achieved higher sensitivity, but used substantially more trainable parameters.
- Pre-training used 94,323 frontal-view chest X-ray images from an NIH-derived dataset covering common thorax diseases and grouped them into five categories.
5. Conclusion
The paper concludes that COVID-CAPS provides effective COVID-19 X-ray diagnosis with few trainable parameters, while pre-training further improves several reported metrics. It also presents comparative results and the external dataset used for pre-training.
- COVID-CAPS is a Capsule Network-based COVID-19 X-ray diagnosis framework with a loss function modified for class imbalance.
- Table 1 reports results for COVID-CAPS alongside results from Reference [11].
- Table 2 describes the external X-ray image dataset used to pre-train COVID-CAPS.
- Pre-training improved COVID-CAPS accuracy, specificity, and AUC, while the framework maintained satisfying performance with few trainable parameters.