Source-linked AI summary
COVID_MTNet: COVID-19 Detection with Multi-Task Deep Learning Approaches
Md Zahangir Alom, M M Shaifur Rahman, Mst Shamima Nasrin, Tarek M. Taha, Vijayan K. Asari
TL;DR
COVID-19 detection is important but can be delayed by diagnostic constraints, motivating faster image-based assessment. The paper combines IRRCNN-based classification with NABLA-N segmentation and evaluates the system on X-ray and CT images. It reports 84.67% X-ray and 98.78% CT testing accuracy, alongside infected-region localization results, while noting limited training samples and false positives in CT segmentation.
Problem
COVID-19 requires timely detection, but RT-PCR diagnosis can be delayed by limited supplies and strict laboratory requirements.
Method
The paper uses IRRCNN for COVID-19 detection and NABLA-N with image-processing steps for infected-region localization in X-ray and CT images.
Results
The system reports 84.67% testing accuracy for X-ray detection and 98.78% for CT detection, with qualitative evidence of infected-region segmentation.
Takeaways & Limitations
The proposed end-to-end system provides COVID-19 detection and infected-region localization across two medical-imaging modalities.
Takeaways & Limitations
The authors state that more COVID-19 samples are needed, especially for the CT model trained and tested on only 300 samples, and that CT segmentation produces some false positives because labeled lung-segmentation data are scarce.
Abstract
from arXiv · showhide
COVID-19 is currently one the most life-threatening problems around the world. The fast and accurate detection of the COVID-19 infection is essential to identify, take better decisions and ensure treatment for the patients which will help save their lives. In this paper, we propose a fast and efficient way to identify COVID-19 patients with multi-task deep learning (DL) methods. Both X-ray and CT scan images are considered to evaluate the proposed technique. We employ our Inception Residual Recurrent Convolutional Neural Network with Transfer Learning (TL) approach for COVID-19 detection and our NABLA-N network model for segmenting the regions infected by COVID-19. The detection model shows around 84.67% testing accuracy from X-ray images and 98.78% accuracy in CT-images. A novel quantitative analysis strategy is also proposed in this paper to determine the percentage of infected regions in X-ray and CT images. The qualitative and quantitative results demonstrate promising results for COVID-19 detection and infected region localization.
1. Introduction
COVID-19 creates an urgent need for timely detection, while diagnostic constraints and rising infections complicate accurate assessment. The paper presents an end-to-end deep-learning system evaluated on X-ray and CT images for detection and infected-region localization.
- Motivation: COVID-19 is a rapidly spreading infectious disease causing respiratory illness and substantial global mortality.The paper describes the disease as newly identified and associated with symptoms ranging from cough and fever to breathing difficulty.
- Motivation: RT-PCR diagnosis can be delayed by limited supplies and strict laboratory requirements.The paper identifies CT image analysis as a faster and easier approach in clinical practice.
- Motivation: Early COVID-19 detection matters for patient prognosis, epidemic control, and public health security.The paper links this need to increasing numbers of affected regions and patients.
- Contribution: The proposed end-to-end system performs COVID-19 detection and infected-region localization using deep learning on X-ray and CT images.Training and validation use publicly available samples collected from different sources worldwide.
- Contribution: The paper evaluates the proposed methods on both imaging modalities and reports promising detection and localization results.The paper is organized around related work, methodology, experiments, results, limitations, and conclusions.
2. Related works
Prior work established deep-learning analysis of X-ray and CT images as a promising direction for COVID-19 detection. The paper builds on this literature by proposing an end-to-end system that detects COVID-19 and localizes infected regions across both modalities.
- Existing approaches: Existing studies commonly use deep-learning methods with X-ray and CT images for COVID-19 detection.The reviewed literature reports promising detection accuracy across these imaging modalities.
- Imaging modalities: CT is described as more sensitive than chest radiography for some pulmonary manifestations, while X-ray remains a common primary screening method.The literature recommends CT particularly when disease is not predominant on X-ray or when assessing lung involvement.
- Representative results: Reported CT-based systems include a model achieving 0.959 ROC AUC with 0.907 sensitivity and 0.911 specificity.That system used pretrained UNet lung segmentation followed by a 3D-CNN trained on 499 CT volumes and tested on 131.
- Representative results: Another CT method grouped COVID-19, Influenza-A viral pneumonia, and healthy cases and reported around 86.7% testing accuracy.The approach used VNet-based region-of-interest segmentation for pulmonary tuberculosis.
- Paper contribution: This paper proposes an end-to-end deep-learning system for COVID-19 detection and infected-region localization from both X-ray and CT images.The proposal follows the literature’s use of both modalities while combining detection with localization.
3. Methodology
The methodology combines separate classification and segmentation models across X-ray and CT workflows. IRRCNN performs COVID-19 detection, while NABLA-N and subsequent image-processing steps identify infected regions.
- Model roles: IRRCNN is used for COVID-19 classification, while NABLA-N performs infected-region segmentation from X-ray and CT images.The methodology assigns classification and region-of-interest detection to separate models.
- COVID-Det_X-ray: For X-ray detection, IRRCNN first classifies normal versus pneumonia images before transfer learning distinguishes COVID-19 from normal images.The pipeline includes data augmentation, model training, and classification outputs for normal and COVID-19 cases.
- COVID-Seg_X-ray: For X-ray segmentation, NABLA-N extracts chest regions, followed by morphological refinement and processing to generate infected-region heatmaps.Chest masks isolate the relevant region before classical image processing and adaptive thresholding identify infection features.
- COVID-Seg_CT: For CT segmentation, NABLA-N is trained on public 2D lung images and applied to CT scans before classical image processing extracts infected regions.The CT workflow uses the same processing pipeline demonstrated for X-ray images.
- Architecture: IRRCNN contains five recurrent residual units, global average pooling, and softmax, while NABLA-N uses encoding and decoding units.The IRRCNN model uses approximately 34M parameters and 1×1 and 3×3 kernels.
4. Experiments
The experiments evaluate multi-model COVID-19 detection and segmentation across X-ray and CT data, using resized images, data augmentation, and separate training configurations.
- Experimental setup: The system uses multiple models for classification and segmentation, implemented with TensorFlow and tested on a single GPU.Training used four NVIDIA GTX2080 Ti GPUs.
- X-ray data: 5,216 pneumonia-detection images comprise 1,341 normal and 3,875 pneumonia samples, resized to 128×128 pixels with class-specific augmentation.
- X-ray data: 704 chest X-ray images with corresponding masks support chest-region segmentation; images were resized to 192×192×3, with 80% used for training and 20% for validation and testing.The resizing substantially reduces the original image information.
- CT data: 420 CT samples include 247 normal and 178 COVID-19 images, resized to 192×192 pixels; 375 were used for training and validation and 45 for testing.The original sample sizes range from 450×338 to 630×630 pixels, and training used data augmentation.
- CT data: 267 labeled 2D CT scans with masks support lung segmentation after resizing from 512×512 to 256×256 single-channel images.Eighty percent of the images were used for training and the remaining 20% for validation and testing.
- Training: The pneumonia-detection model used Adam with learning rate 1 × 10^-3, batch size 32, and 75 epochs with scheduled learning-rate reductions.The learning rate was reduced by a factor of 10 after each 25 epochs.
5. Results
The proposed models detect COVID-19 and localize infected regions across X-ray and CT images using classification and segmentation pipelines. Results include quantitative detection, segmentation, and infection-percentage estimates, with some false detections observed in COVID-19 CT samples.
- COVID-Det_X-ray: 84.67% testing accuracy was achieved for COVID-19 detection on 67 new X-ray samples using pretrained pneumonia-model weights.The same IRRCNN model first achieved 87.26% testing accuracy for pneumonia detection on 624 new images.
- COVID-Det_X-ray: 0.93 AUC was reported for the pneumonia detection model.
- COVID-Seg_X-ray: 0.8650 IoU and 0.8846 Dice similarity coefficient were reported for the X-ray chest-segmentation model.The COVID-Seg_X-ray pipeline refined chest masks, extracted chest regions, and identified infected regions in the resulting images.
- COVID-Seg_X-ray: 33.52% and 37.58% infection percentages were calculated for two X-ray examples relative to total lung-region pixels.The analysis also reported approximately 51.23%, 27.66%, and 47.89% for three additional examples and was proposed as a measure of disease severity.
- Abdominal CT evaluation: The X-ray segmentation workflow was also evaluated on abdominal CT images using refined masks, extracted lung regions, adaptive thresholding, and infected-region heatmaps.
- COVID-Seg_CT: 0.9885 F1-score and 0.9956 global accuracy were obtained for COVID-Seg_CT on new lung-segmentation samples.The model was applied to COVID-19 CT samples without labeled COVID-19 segmentation data; false detection occurred in some cases.
6. Discussions
The discussion reports stronger pneumonia accuracy for the proposed IRRCNN model than a cited published result and emphasizes pixel-level infected-region segmentation over patch-based detection. It presents this strategy as reducing the possibility of false positive and false negative detections within lung regions.
- Infected-region localization: Pixel-level infected-region segmentation was proposed instead of patch-based extraction to address possible false positive and false negative detections.The authors state that qualitative experiments demonstrated efficient infected-region detection within the lung part.
7. Limitations
The study reports strong detection and segmentation results but identifies data limitations that constrain robustness and segmentation quality. More COVID-19 samples and labeled CT segmentation data are needed for improvement.
- The COVID-Det_X-ray model needs more COVID-19 samples for training and testing.
- The COVID-Det_CT model was trained and tested on only 300 samples, limiting its generalization and robustness.
- Scarce labeled CT lung-segmentation samples lead COVID-Seg_CT to produce some false-positive detections.The paper identifies false positives in the second row of Figure 15(d) as requiring improvement.
8. Conclusion
The paper proposes an end-to-end system for COVID-19 detection and infected-region localization across multiple medical-imaging modalities. Its models achieve 84.67% testing accuracy on X-ray images and 98.78% on CT images, while qualitative results show accurate infected-region segmentation.
- The study proposes an end-to-end system for COVID-19 detection and infected-region localization from two medical-imaging modalities.
- The improved IRRCNN and NABLA-3 models are applied to classification and segmentation tasks on publicly available X-ray and CT datasets.
- 84.67% testing accuracy is achieved for COVID-19 detection from X-ray images, compared with 98.78% from CT images.
- Qualitative results demonstrate accurate segmentation and detection of COVID-19 infected regions in both X-ray and CT images.
- The authors plan to collect more COVID-19 samples to develop a more robust and accurate system.