Source-linked AI summary
Lung Infection Quantification of COVID-19 in CT Images with Deep Learning
Fei Shan, Yaozong Gao, Jun Wang, Weiya Shi, Nannan Shi, Miaofei Han, Zhong Xue, Dinggang Shen, Yuxin Shi
TL;DR
Radiologists lack computerized tools to accurately quantify COVID-19 severity from CT images. The paper develops a deep learning segmentation system assisted by human-in-the-loop training, with POI showing clinical relevance to pneumonia severity and enabling infection-distribution analysis.
Problem
Radiologists lack a computerized tool to accurately quantify COVID-19 severity, including the percentage of infection in the whole lung.
Method
The paper develops a deep learning system for segmenting COVID-19 infection regions and uses human-in-the-loop training, where radiologists iteratively correct model outputs to update the network.
Results
A Pearson correlation coefficient of 0.5 between CT-derived whole-lung POI and PSI in 196 patients indicated moderate correlation with pneumonia severity.
Takeaways & Limitations
The system supports quantitative analysis of infection severity and distribution across lung lobes and bronchopulmonary segments.
Takeaways & Limitations
Validation used data from one center, so generalization to COVID-19 patients from other geographic areas requires further multi-center validation.
Abstract
from arXiv · showhide
CT imaging is crucial for diagnosis, assessment and staging COVID-19 infection. Follow-up scans every 3-5 days are often recommended for disease progression. It has been reported that bilateral and peripheral ground glass opacification (GGO) with or without consolidation are predominant CT findings in COVID-19 patients. However, due to lack of computerized quantification tools, only qualitative impression and rough description of infected areas are currently used in radiological reports. In this paper, a deep learning (DL)-based segmentation system is developed to automatically quantify infection regions of interest (ROIs) and their volumetric ratios w.r.t. the lung. The performance of the system was evaluated by comparing the automatically segmented infection regions with the manually-delineated ones on 300 chest CT scans of 300 COVID-19 patients. For fast manual delineation of training samples and possible manual intervention of automatic results, a human-in-the-loop (HITL) strategy has been adopted to assist radiologists for infection region segmentation, which dramatically reduced the total segmentation time to 4 minutes after 3 iterations of model updating. The average Dice simiarility coefficient showed 91.6% agreement between automatic and manual infaction segmentations, and the mean estimation error of percentage of infection (POI) was 0.3% for the whole lung. Finally, possible applications, including but not limited to analysis of follow-up CT scans and infection distributions in the lobes and segments correlated with clinical findings, were discussed.
Abstract
The paper concerns COVID-19 infection-region segmentation and quantification using CT, deep learning, and human-in-the-loop methods.
- The study focuses on COVID-19 infection region segmentation in CT images.
- Deep learning is used as the paper’s core computational approach.
- Human-in-the-loop interaction is included to support infection-region analysis.
Introduction
CT is useful for COVID-19 screening and follow-up, but infection quantification remains limited by the absence of computerized tools and laborious manual contouring. The paper therefore develops a deep learning system with human-in-the-loop training to automate quantitative assessment.
- Chest CT is used as a complementary COVID-19 screening indicator because scans frequently show bilateral patchy shadows or ground-glass opacity.
- Follow-up CT scans every 3-5 days are recommended to evaluate therapeutic responses during rapid disease progression.
- Radiological reports remain largely qualitative because computerized tools for quantifying infection regions and longitudinal changes are lacking.
- Manual infection-region contouring is tedious, time-consuming, and potentially inconsistent, motivating a fast automatic contouring tool.
- The proposed system uses deep learning for infection segmentation and human-in-the-loop correction to iteratively generate training samples.
Materials and Methods
The study trains and evaluates a VB-Net-based CT segmentation pipeline using retrospective COVID-19 datasets and radiologist-assisted iterative annotation. The system computes infection volumes and percentage-of-infection measures across lung regions, with DSC and Pearson correlation used for evaluation.
- Datasets: Validation used 300 CT scans from 300 COVID-19 patients, while 249 scans from other centers were used for training.
- DL-Based Segmentation Network: VB-Net: VB-Net is a modified 3-D convolutional network designed to segment difficult, low-contrast infection regions with variable shapes and positions.
- DL-Based Segmentation Network: VB-Net: A bottleneck architecture reduces model size and inference time while fusing cross-channel features for large 3D volumetric data.
- Training VB-Net with Human-In-The-Loop Strategy: Human-in-the-loop training has radiologists correct model outputs from successive data batches, adding corrected results as training data for model updates.
- Quantitative Assessment: The pipeline computes infection volumes and POIs for the whole lung, lobes, and bronchopulmonary segments to quantify severity and distribution.
- Statistical Analysis and Evaluation Metrics: The Dice similarity coefficient measures overlap between automatic segmentation S and radiologist reference region R, while Pearson correlation evaluates association between two variables.
Results
The deep learning system produced infection segmentations and percentage-of-infection estimates that closely matched radiologist assessments. Its measurement errors were near inter-rater variability, while human-in-the-loop training reduced annotation time across model updates.
- Segmentation Performance: 91.6% average Dice similarity coefficient indicated agreement between automatic and manual infection segmentations.
- POI Estimation: 0.3% mean POI estimation error was observed for the whole lung, compared with 0.5% for lobes and 0.8% for bronchopulmonary segments.
- Inter-Rater Variability: Radiologist inter-rater comparison produced a 96.1%±3.5% average Dice similarity coefficient and a 7.4±5.2 cm^3 average volume measurement difference.
- Comparison with Radiologists: The deep learning system’s segmentation and measurement errors were close to inter-rater variability.
- Human-in-the-Loop Evaluation: Human-in-the-loop evaluation compared manual labeling time and segmentation accuracy across successive model-training stages.
Discussion
The system combines deep-learning infection segmentation with radiologist-guided training to quantify COVID-19 burden and support longitudinal and distributional analysis. It showed clinically relevant severity association while remaining limited by single-center validation and infection-specific scope.
- The system segments COVID-19 infection regions on CT and quantifies their shapes, volumes, and percentage of infection.
- Follow-up CT scans every 3–5 days can be analyzed for longitudinal changes in infection volume, consolidation, and GGO.Surface rendering can visualize these changes after infection-region segmentation.
- 0.5 Pearson correlation coefficient was observed between whole-lung POI and PSI in 196 patients, indicating a moderate association with pneumonia severity.The authors describe this result as clinically relevant and note ongoing work on progression prediction.
- Mean POIs were higher in the left and right lower lobes than in other lobes across 300 CT scans.The system also enables large-scale quantitative analysis of lobe and bronchopulmonary-segment distributions.
- The most frequently infected listed segments included right lower-lobe outer basal, dorsal, and posterior basal segments, followed by left lower-lobe outer basal and dorsal segments.
- HITL training lets radiologists iteratively intervene in model outputs, accelerating annotation while incorporating professional knowledge.The strategy also makes model performance more comprehensible to radiologists during training.
- The validation data came from one center, and the system was developed to quantify COVID-19 infection rather than other pneumonias.The authors identify multi-center validation and extension to other pneumonia as future work.
Figures
The figures present the segmentation network, human-in-the-loop workflow, quantitative assessment pipeline, representative disease-stage results, follow-up progression, and regional POI distributions.
- Figure 1 depicts the COVID-19 infection-segmentation network, with dashed boxes identifying bottle-neck structures inside its V-shaped architecture.
- Figure 2 presents the human-in-the-loop workflow for developing and using the segmentation system.
- Figure 3 shows CT input passing through deep-learning segmentation before infection volumes and POIs are calculated for the whole lung, lobes, and bronchopulmonary segments.
- Figure 4 compares early, progressive, and severe cases using CT images, segmentation overlays, and 3D renderings of segmented infections.
- Figure 5 follows one patient’s POI progression and gradual recovery, distinguishing ground-glass opacities in green from consolidation in red.
- Figure 6 displays POI distributions across 5 lung lobes and 18 bronchopulmonary segments using box-and-whisker plots.