| Filename | Latest commit message | Latest commit date |
|---|---|---|
| figures | ||
| README.md | ||
| Trained_MDA_SN_Model.mat | ||
Meta-Domain Adaptive Framework for Efficient Diagnostic Assessment of Lung Infection Using CT Radiographs
Lightweight AI framework for lung infection segmentation, infection ratio quantification, and semantic retrieval of clinically similar CT cases.
🌟 Overview
MDA-SN (Meta-Domain Adoptive Segmentation Network) is a lightweight and computationally efficient AI framework developed for automated lung infection assessment from chest CT scans. The framework combines adaptive domain normalization, attention-driven segmentation, quantitative infection analysis, and semantic retrieval to provide an end-to-end computer-aided diagnostic solution.
🎯 Key Capabilities
| Capability | Description |
|---|---|
| 🫁 Lung Segmentation | Automatic extraction of lung regions from CT scans |
| 🔥 Infection Segmentation | Precise localization of infection regions |
| 📊 Infection Ratio Quantification | Objective estimation of infection burden |
| 🔍 Semantic Case Retrieval | Retrieval of clinically similar CT cases |
| 📈 Disease Severity Assessment | Quantitative evaluation of infection progression |
| 🌍 Cross-Dataset Generalization | Robust performance across diverse CT datasets |
🚀 Why MDA-SN?
- Lightweight architecture with only 6.22M parameters
- Real-time processing at approximately 29 CT slices/second
- Adaptive Data Normalization (ADN) for improved cross-domain robustness
- Multi-scale semantic learning using grouped and dilated convolutions
- Semantic attention-driven retrieval framework for case-based analysis
- Designed for explainable and reproducible medical AI research
🏥 Clinical Impact
The proposed framework enables automated lung infection screening, infection severity quantification, and retrieval of clinically relevant CT cases, supporting radiologists and healthcare professionals in diagnostic assessment, treatment planning, and patient monitoring.
Graphical Abstract
High-Level Workflow
Input Chest CT Scan
│
▼
Adaptive Data Normalization
│
▼
MDA-SN Segmentation Network
│
├── Lung Region Segmentation
│
└── Infection Region Segmentation
│
▼
Infection Ratio Calculation
│
▼
Dice-Based Similarity Matching
│
▼
Top-N Similar CT Case Retrieval
│
▼
Clinical Decision Support
Key Contributions
| No. | Contribution |
|---|---|
| 1 | A lightweight Meta-Domain Adoptive Segmentation Network for lung infection segmentation in CT scans. |
| 2 | Adaptive Data Normalization strategy to reduce domain shift across CT datasets. |
| 3 | Multi-scale dilated grouped convolution with residual attention for efficient lesion localization. |
| 4 | Infection Ratio calculation for quantitative infection severity assessment. |
| 5 | Semantic attention-driven retrieval framework for retrieving clinically similar CT slices. |
| 6 | Real-time performance with approximately 29 slices per second and only 6.22M parameters. |
Method Summary
1. Adaptive Data Normalization
The proposed ADN strategy aligns the visual distribution of testing CT slices with the training domain to improve cross-dataset generalization.
2. MDA-SN Segmentation
MDA-SN uses:
- Residual blocks
- Stride blocks
- Atrous/dilated convolution block
- Residual attention
- Lightweight decoder
3. Infection Ratio Quantification
The infection burden is computed as:
IR = Infected Area / Lung Area × 100
4. Similar CT Case Retrieval
The predicted infection mask of the query CT slice is compared with database masks using Dice similarity.
Dice = 2 × |Query Mask ∩ Retrieval Mask| / (|Query Mask| + |Retrieval Mask|)
The top-ranked CT slices are retrieved as clinically similar cases.
Datasets
The proposed framework was evaluated on two publicly available CT datasets.
| Dataset | Description | Annotation |
|---|---|---|
| COVID-19-CT-Seg | CT slices from COVID-19 patients | Lung and infection masks |
| MosMed | COVID-19 CT scans with infection annotations | Infection masks |
Experimental Setup
| Item | Details |
|---|---|
| Software | MATLAB R2020b |
| GPU | NVIDIA GeForce GTX 1070 |
| CPU | Intel Core i7 |
| RAM | 16 GB |
| Optimizer | Stochastic Gradient Descent |
| Initial Learning Rate | 0.001 |
| Evaluation Metrics | SEN, SPE, Dice, IoU |
| Validation | Five-fold cross-validation |
Main Results
Within-Dataset Performance
| Experiment | Dataset | SEN (%) | SPE (%) | Dice (%) | IoU (%) |
|---|---|---|---|---|---|
| Experiment 1 | COVID-19-CT-Seg | 92.46 | 99.12 | 83.37 | 75.01 |
| Experiment 2 | MosMed | 89.65 | 99.41 | 69.18 | 61.81 |
Cross-Dataset Performance
| Setting | SEN (%) | SPE (%) | Dice (%) | IoU (%) |
|---|---|---|---|---|
| Without ADN | 63.68 | 99.34 | 71.23 | 63.22 |
| With ADN | 67.91 | 99.64 | 75.93 | 67.42 |
Computational Efficiency
| Method | Parameters | Speed | FLOPs |
|---|---|---|---|
| Proposed MDA-SN | 6.22M | 29.11 slices/s | 28.26 |
Visual Results
Retrieval Results
Applications
| Application | Description |
|---|---|
| Lung infection assessment | Automatic localization of infection regions in CT scans. |
| Infection severity estimation | Quantitative infection ratio calculation. |
| Clinical monitoring | Tracking infection burden over time. |
| Case-based retrieval | Retrieval of similar CT cases from the database. |
| Decision support | Supporting clinicians with objective AI-based analysis. |
Algorithmic Workflow
The proposed MDA-SN framework consists of three main phases: training, testing, and retrieval.
Step 1: Training Phase
Input:
- Training CT images
- Corresponding lung and infection masks
Process:
1. Load training CT scans and ground-truth masks.
2. Initialize the MDA-SN model.
3. Train one model for lung region segmentation.
4. Train another model for infection region segmentation.
5. Optimize the model using weighted segmentation loss.
6. Save the trained MDA-SN models.
Output:
- Trained lung segmentation model
- Trained infection segmentation model
Step 2: Testing Phase
Input:
- Test CT scan
- Trained MDA-SN models
Process:
1. Apply adaptive data normalization to reduce domain variation.
2. Generate the lung segmentation mask.
3. Generate the infection segmentation mask.
4. Remove non-lung regions from the infection mask.
5. Produce the final infection prediction map.
Output:
- Predicted lung mask
- Predicted infection mask
Step 3: Infection Ratio Quantification
Input:
- Predicted lung mask
- Predicted infection mask
Process:
1. Count the total number of lung pixels.
2. Count the total number of infected pixels.
3. Calculate infection ratio using:
Infection Ratio = (Infected Area / Lung Area) × 100
Output:
- Infection Ratio percentage
Step 4: Semantic Retrieval Phase
Input:
- Query infection mask
- Database of CT masks
Process:
1. Compare the query mask with each database mask.
2. Calculate Dice similarity score for each comparison.
3. Rank all database CT slices based on Dice similarity.
4. Select the Top-N most similar CT cases.
5. Retrieve corresponding CT images and infection ratios.
Output:
- Top-N clinically similar CT cases
- Dice similarity scores
- Infection ratio values
Step 5: Clinical Output
Final Output:
- Segmented lung region
- Segmented infection region
- Infection ratio score
- Retrieved similar CT cases
- Case-based support for diagnostic assessment
Paper Information
Title:
Meta-Domain Adaptive Framework for Efficient Diagnostic Assessment of Lung Infection Using CT Radiographs
Authors:
Muhammad Owais, Taimur Hassan, Naqash Afzal, Saddam Hussain Khan, Divya Velayudhan, Iyyakutti Iyappan Ganapathi, Irfan Hussain, and Naoufel Werghi
Corresponding Author:
Muhammad Owais
Khalifa University Center for Autonomous Robotic Systems, Khalifa University, Abu Dhabi, UAE
Email: muhammad.owais@ku.ac.ae
Citation
@article{owais2026mdasn,
title={Meta-Domain Adaptive Framework for Efficient Diagnostic Assessment of Lung Infection Using CT Radiographs},
author={Owais, Muhammad and Hassan, Taimur and Afzal, Naqash and Khan, Saddam Hussain and Velayudhan, Divya and Ganapathi, Iyyakutti Iyappan and Hussain, Irfan and Werghi, Naoufel},
journal={Scientific Reports},
year={2026}
}
License
This repository is released for academic and research purposes.
Contact
For questions, please contact:
Muhammad Owais
Khalifa University, Abu Dhabi, UAE
Email: muhammad.owais@ku.ac.ae
GitHub: https://github.com/Owais-CodeHub