Research map / Image generation / GANs
COPD-FlowNet: Elevating Non-invasive COPD Diagnosis with CFD Simulations (Student Abstract)
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes COPD-FlowNet, a custom GAN that augments a small set of CFD tracheal velocity-field images from COPD patients, plus a custom CNN that classifies obstruction location (left, right, both).
Problem
Non-invasive diagnosis of COPD obstruction location from CFD airflow simulations is limited by the cost of generating CFD simulations and by class imbalance in a small dataset of 295 tracheal velocity-field images.
Approach
Use a custom GAN to synthesize class-balanced 128x128 CFD velocity field images (raising the set to 1200 images, 400 per class) and train a custom shallow CNN classifier on the augmented data to predict obstruction location.
Key results
- Custom CNN accuracy rose from 73.94% before augmentation to 96.35% after GAN augmentation.
- Reported 96.35% accuracy versus 96% for ResNet-50 and 95.00% for EfficientNet on the same task, a small margin.
- Dataset expanded from 295 to 1200 images, balanced at 400 per class (left, right, both).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply augmentation techniques to the classification dataset | COPD-FlowNet CNN · COPD obstruction classification | accuracy: 73.94% -> 96.35% | Helped |
| Use COPD-FlowNet proposed CNN instead of ResNet-50 | COPD obstruction classification · COPD obstruction classification | accuracy: 96% -> 96.35% | Helped |
| Use COPD-FlowNet proposed CNN instead of EfficientNet | COPD obstruction classification · COPD obstruction classification | accuracy: 95.00% -> 96.35% | Helped |
Limitations
- Classifies only three coarse locations (left, right, both lungs), not specific obstruction sites.
- Small dataset of 295 original images; the GAN's realism and diversity are not quantitatively evaluated.
- Comparison with ResNet-50 and EfficientNet has small margins and no reported variance or significance testing.
Experimental setup
Datasets: CFD velocity flow field images of COPD patient tracheas (295 images, from Hu et al. 2022)
Benchmarks: COPD CFD velocity-field classification dataset
Metrics: accuracy precision recall F1 score ROC curve
Baselines: ResNet-50 EfficientNet
Builds on: GAN
Closest papers by meaning
- Learning Low-Rank Feature for Thorax Disease Classification (NeurIPS 2024 · similarity 0.77)
- Finite Volume Features, Global Geometry Representations, and Residual Training for Deep Learning-based CFD Simulation (ICML 2024 · similarity 0.77)
- MS$^3$D: A RG Flow-Based Regularization for GAN Training with Limited Data (ICML 2024 · similarity 0.77)
- Resp-Agent: An Agent-Based System for Multimodal Respiratory Sound Generation and Disease Diagnosis (ICLR 2026 · similarity 0.77)
- MeshMask: Physics-Based Simulations with Masked Graph Neural Networks (ICLR 2025 · similarity 0.77)
- TCAM-Diff: Triplane-Aware Cross-Attention Medical Diffusion Model (AAAI 2025 · similarity 0.76)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.