[Paper Review] Radiologist-Level COVID-19 Detection Using CT Scans with Detail-Oriented Capsule Networks

Aryan Mobiny, Pietro Antonio Cicalese+7 more|arXiv (Cornell University)|Apr 16, 2020

COVID-19 diagnosis using AI38 references73 citations

TL;DR

The paper introduces DECAPS, a detail-oriented capsule network with inverted dynamic routing, Peekaboo training, and pix2pix data augmentation to detect COVID-19 from CT scans, achieving radiologist-level performance and outperforming several baselines.

ABSTRACT

Radiographic images offer an alternative method for the rapid screening and monitoring of Coronavirus Disease 2019 (COVID-19) patients. This approach is limited by the shortage of radiology experts who can provide a timely interpretation of these images. Motivated by this challenge, our paper proposes a novel learning architecture, called Detail-Oriented Capsule Networks (DECAPS), for the automatic diagnosis of COVID-19 from Computed Tomography (CT) scans. Our network combines the strength of Capsule Networks with several architecture improvements meant to boost classification accuracies. First, DECAPS uses an Inverted Dynamic Routing mechanism which increases model stability by preventing the passage of information from non-descriptive regions. Second, DECAPS employs a Peekaboo training procedure which uses a two-stage patch crop and drop strategy to encourage the network to generate activation maps for every target concept. The network then uses the activation maps to focus on regions of interest and combines both coarse and fine-grained representations of the data. Finally, we use a data augmentation method based on conditional generative adversarial networks to deal with the issue of data scarcity. Our model achieves 84.3% precision, 91.5% recall, and 96.1% area under the ROC curve, significantly outperforming state-of-the-art methods. We compare the performance of the DECAPS model with three experienced, well-trained thoracic radiologists and show that the architecture significantly outperforms them. While further studies on larger datasets are required to confirm this finding, our results imply that architectures like DECAPS can be used to assist radiologists in the CT scan mediated diagnosis of COVID-19.

Motivation & Objective

• Motivate rapid, accurate COVID-19 screening when radiology expertise is limited.
• Develop a model that can focus on informative regions while leveraging context.
• Address data scarcity with synthetic data augmentation and targeted training strategies.
• Benchmark against experienced radiologists to assess practical clinical utility.

Proposed method

• Propose DECAPS, a Capsule Network architecture with two key ideas: detail-oriented routing to emphasize ROIs and an inverted dynamic routing mechanism.
• Introduce IDR to have lower-level capsules compete for attention of higher-level capsules, yielding capsule heads and activation maps per class.
• Implement a Peekaboo training regime that crops/drops patches guided by activation maps to encourage learning from multiple ROIs.
• Augment data using a pix2pix conditional GAN with PatchGAN and U-net structure to generate realistic COVID-19 CT samples.
• Use a two-stage training/prediction strategy including coarse predictions from full images and fine-grained predictions from ROIs, followed by distillation to final predictions.

Experimental results

Research questions

• RQ1 Can a capsule-based network outperform traditional CNNs on CT-based COVID-19 detection with limited data?
• RQ2 Do detailed ROI-focused routing and training strategies improve localization and classification performance?
• RQ3 What is the impact of data augmentation with GAN-generated CT samples on model generalization?
• RQ4 How does the model compare to expert radiologists in diagnostic accuracy?

Key findings

• DECAPS with Peekaboo outperformed baseline CNNs and achieved the highest AUC of 0.961±0.009.
• DECAPS alone reached AUC 0.927±0.017, outperforming Inception-v3, DenseNet121, and ResNet50 on the same dataset.
• The combined DECAPS+Peekaboo with pix2pix augmentation yielded the best performance (AUC 0.961±0.009).
• The model demonstrated higher recall (0.915±0.057) and comparable precision (0.843±0.024) relative to other architectures.
• Compared to three radiologists, the architecture significantly outperformed human experts in the Alpha-Test set performance.
• Activation maps per capsule head enable weakly-supervised localization of regions contributing to the decision.