Weakly Supervised Deep Learning for COVID-19 Infection Detection and Classification from CT Images (2004.06689v1)

Abstract: An outbreak of a novel coronavirus disease (i.e., COVID-19) has been recorded in Wuhan, China since late December 2019, which subsequently became pandemic around the world. Although COVID-19 is an acutely treated disease, it can also be fatal with a risk of fatality of 4.03% in China and the highest of 13.04% in Algeria and 12.67% Italy (as of 8th April 2020). The onset of serious illness may result in death as a consequence of substantial alveolar damage and progressive respiratory failure. Although laboratory testing, e.g., using reverse transcription polymerase chain reaction (RT-PCR), is the golden standard for clinical diagnosis, the tests may produce false negatives. Moreover, under the pandemic situation, shortage of RT-PCR testing resources may also delay the following clinical decision and treatment. Under such circumstances, chest CT imaging has become a valuable tool for both diagnosis and prognosis of COVID-19 patients. In this study, we propose a weakly supervised deep learning strategy for detecting and classifying COVID-19 infection from CT images. The proposed method can minimise the requirements of manual labelling of CT images but still be able to obtain accurate infection detection and distinguish COVID-19 from non-COVID-19 cases. Based on the promising results obtained qualitatively and quantitatively, we can envisage a wide deployment of our developed technique in large-scale clinical studies.

• The paper presents a weakly supervised deep learning framework that reduces manual annotation while achieving a high AUC (up to 0.97) in COVID-19 CT scan classification.
• It utilizes a multi-scale VGG-inspired CNN with saliency mapping and GMP-based aggregation for robust feature extraction from diverse CT scans.
• The framework offers a scalable, clinical tool that discriminates COVID-19 from CAP and NP, complementing RT-PCR tests with improved efficiency.

Weakly Supervised Deep Learning for COVID-19 Detection and Classification from CT Images: An Overview

This paper focuses on a methodological advancement within the domain of medical imaging, specifically on the application of weakly supervised deep learning techniques for the detection and classification of COVID-19 infection using CT images. The novelty of the proposed approach lies in its ability to effectively minimize the requirement for manual labeling while retaining accuracy in detection and classification tasks, particularly distinguishing COVID-19 cases from community-acquired pneumonia (CAP) and non-pneumonia (NP) cases.

The paper utilizes a retrospective data collection from 450 CT scans across multiple centers and scanners obtained between September 2016 and March 2020, accommodating a diverse set of imaging conditions. This diversity is essential to ensure that the model trained is robust enough to handle variations inherent in clinical imaging.

Methodology and Framework

The paper introduces a weakly supervised deep learning framework that leverages multi-scale learning and saliency maps to tackle the inherent challenges in CT image classification. A striking feature of the methodology is the use of a VGG-inspired CNN architecture to perform multi-scale feature extraction, which is further enhanced by a Global Max Pooling (GMP) based multi-level aggregation strategy.

The framework employs a lung segmentation model trained on an external dataset, the TCIA dataset, using a multi-view U-Net. This approach standardizes lung segmentation, facilitating more accurate downstream infection detection and classification tasks. The large-scale convolutional network is trained with a focal loss to handle class imbalance effectively, which is critical given the varied nature of the dataset in terms of class distribution.

The paper also demonstrates the utility of Integrated Gradients to generate high-resolution saliency maps, allowing for precise lesion localization beyond simple classification. The proposed method notably achieves strong results with an area under the curve (AUC) of up to 0.97 for NP and COVID-19 discrimination, attesting to its robust performance.

Statistical Evaluation

In-depth statistical evaluations reported a classification accuracy of 87.4% overall for the joint classifier in three-way tasks and significantly higher accuracies in the binary classification scenarios, particularly in distinguishing COVID-19 from NP cases. The paper extensively contrasts its results with a re-implementation of the Navigator-Teacher-Scrutinizer Network (NTS-NET), showcasing improvements in performance metrics across various tasks.

Implications and Future Directions

Practically, this framework offers a scalable solution that can reduce the burden of manual annotation in clinical environments, increasingly vital during pandemic conditions when resources are stretched. The framework is proposed as an auxiliary tool, complementing RT-PCR tests whose sensitivity may be suboptimal due to false negatives, especially when rapid decision-making is crucial.

Theoretically, this work could pave the way for further exploration into weakly supervised learning techniques applied in different medical imaging contexts, particularly where data labeling is a significant barrier. The promising performance of this framework suggests a potential for extending such methodologies to a broader range of medical imaging applications.

Future developments could involve extending the core methodology to employ more sophisticated CNN architectures, such as ResNet or Inception, to further increase the network's capacity and potential adaptability to different imaging modalities or conditions.

In conclusion, the paper presents an insightful paper with compelling evidence for the efficacy of weakly supervised deep learning in the nuanced and critical task of COVID-19 detection and classification from CT images, with broad implications for enhancing the efficiency and accuracy of pandemic response strategies.