Source-linked AI summary
Coronavirus Detection and Analysis on Chest CT with Deep Learning
Ophir Gozes, Maayan Frid-Adar, Nimrod Sagie, Huangqi Zhang, Wenbin Ji, Hayit Greenspan
TL;DR
The paper addresses the challenge of data-dependent deep learning when early disease knowledge and annotations are limited. It builds a weakly supervised chest-CT pipeline for COVID-19 detection, localization, severity scoring, and exploratory clustering. The system achieved 0.948 AUC for case-level classification, and the authors report distinct manifestation clusters and a relationship between Corona score and clinical grade.
Problem
Data-dependent deep learning is challenging when samples are scarce, and early disease knowledge limited exact annotation of radiological manifestations.
Source: §1 · pp. 1–2Method
Results
The system achieved 0.948 AUC for Coronavirus versus non-coronavirus cases on Chinese control and infected patients; exploratory clusters corresponded to normal, focal, and diffuse manifestations, and Corona score corresponded to clinical grade.
Source: §4 · pp. 7–8Takeaways & Limitations
Fine-grain localization offers radiologists insight into model predictions and 3D case-level visualization, while the authors hypothesize that Corona score may support patient monitoring and management.
Source: §4 · pp. 7–8
Abstract
from arXiv · showhide
The outbreak of the novel coronavirus, officially declared a global pandemic, has a severe impact on our daily lives. As of this writing there are approximately 197,188 confirmed cases of which 80,881 are in "Mainland China" with 7,949 deaths, a mortality rate of 3.4%. In order to support radiologists in this overwhelming challenge, we develop a deep learning based algorithm that can detect, localize and quantify severity of COVID-19 manifestation from chest CT scans. The algorithm is comprised of a pipeline of image processing algorithms which includes lung segmentation, 2D slice classification and fine grain localization. In order to further understand the manifestations of the disease, we perform unsupervised clustering of abnormal slices. We present our results on a dataset comprised of 110 confirmed COVID-19 patients from Zhejiang province, China.
1 Introduction
The paper addresses deep learning’s data dependence and limited disease knowledge and annotation in the early pandemic. It proposes a chest-CT system to detect suspected COVID-19, localize abnormalities, and estimate severity.
1 Introduction
- Chest CT findings including ground-glass opacities, crazy paving, and increasing consolidation were presented as potential cues for COVID-19 detection and clinical decision-making.PCR is described as the gold standard for confirmation; CT is presented as a potential tool.Source: §1 · p. 1
- Early disease knowledge and radiological annotations were limited, while data-dependent deep learning posed a challenge when data were scarce.The authors describe these constraints as motivating methods that can generalize from relatively few samples with weak supervision.Source: §1 · pp. 1–2
- The proposed approach sequences 2D processing with 3D fusion, allowing use of pretrained architectures to reduce deep-learning training time and resources.The introduction also identifies image clustering, segmentation, and classification as methods for assessing disease progression and guiding care.Source: §1 · p. 2
- For suspected COVID-19 cases, the system processes non-contrast chest CT scans and outputs an abnormality localization map and a disease-severity score.The score is produced for cases classified as positive.Source: §1 · p. 2
2 Methods
The methods form a CT-analysis pipeline: lung segmentation and ROI selection feed slice classification, multi-scale Grad-CAM localization, and case-level severity scoring, with clustering used to explore manifestations.
2 Methods
- The pipeline localizes lungs, classifies 2D ROIs, extracts multi-scale Grad-CAM maps, combines slice detections and maps into a volumetric Corona score, and clusters normal and abnormal slices.The methods use several datasets for development and testing.Source: §2 · pp. 2–3
2.1 Lung Segmentation and lung ROI selection
- Lung segmentation uses an ImageNet-pretrained U-Net trained on 6,150 abnormal-lung CT slices and masks; the largest predicted-mask bounding box defines each lung crop.The crop is intended to focus classification on lung-related regions.Source: §2.1 · p. 3
2.2 COVID-19 Classifier
- The COVID-19 ROI classifier uses an ImageNet-pretrained 2D ResNet-50 [7] with 224×224 inputs, trained on 1,865 annotated slices from suspected cases at several Chinese hospitals; random splits were 1,725/320/270, with no patient overlap between test and development sets.Annotations comprised 1,036 normal and 829 abnormal slices; augmentation included rotations, horizontal flips, and cropping.Source: §2.2 · pp. 3–4
2.3 Model Interpretation and multi-scale GradCam for Localization
- For fine-grain localization, Grad-CAM maps at 14×14 and 28×28 resolutions are normalized and multiplied to form a combined heatmap.The passage describes this as a way to check whether learning focused on pathological areas.Source: §2.3 · p. 4, Figure 2 · p. 4
2.4 Case-wise COVID-19 Scoring for Severity Estimation
- Case-level Corona score sums activation-map values above a predefined threshold across positive slices; experiments set the threshold to 0.6 by expert visual evaluation.The score aggregates the maps into a 3D localization volume representing disease extent through the lungs.Source: §2.4 · p. 4
3 Experiments and Results
The corona score classified COVID-19 versus non-Coronavirus patients at the case level and was evaluated against radiologist severity grades. Unsupervised clustering separated negative slices from positive slices with focal or diffuse patterns.
Corona score for disease detection
- 0.948 AUC (95%CI: 0.912-0.985) was achieved using the corona score to classify Chinese control and infected patients at the case level.The analysis used 109 confirmed COVID-19 patients and 90 non-Coronavirus patients.Source: Corona score for disease detection · p. 5
Corona score as a criteria for severity
- The severe and non-severe groups had median corona scores of 61.5 cm3 and 227.5 cm3, respectively, with p=0.0064 for the two-sided Wilcoxon rank sum test.The comparison used first-time-point scores from 49 patients classified by a radiologist as severe (n=13) or non-severe (n=36).Source: Corona score as a criteria for severity · pp. 5–6, Figure 4 · p. 6
3.3 Analysis of the disease manifestations
- Negative slices were well distinguished from positive slices, which formed groups with subtle focal lesions and diffuse severe patterns.The analysis used positive key slices from 110 patients and negative slices from 81 patients.Source: §3.3 · pp. 6–7, Figure 5 · p. 7
4 Discussion
The discussion presents weakly supervised deep learning for COVID-19 detection, quantification, and exploration on chest CT. It reports case-level classification performance and proposes clinical monitoring as a possible use of the corona score.
4 Discussion
- The study reports weakly supervised deep-learning results for Coronavirus detection, quantification, and exploration on chest CT.The authors describe it as one of the first reports in this domain, to their knowledge.Source: §4 · p. 7
- 0.948 AUC (95%CI: 0.912-0.985) was achieved for Coronavirus versus non-Coronavirus classification per thoracic CT study on Chinese control and infected patients.The authors also describe fine-grain localization as potentially useful to radiologists for inspecting model predictions and viewing cases in 3D.Source: §4 · p. 7
- Exploratory analysis identified clusters corresponding to Normal, Focal, and Diffuse disease manifestations.The authors describe unsupervised clustering as a way to automatically explore a new unknown disease.Source: §4 · p. 8
- The authors report that the corona score corresponds to clinical disease grade and hypothesize it may support patient monitoring and management.This is presented as a hypothesis about potential use.Source: §4 · p. 8