Enhancing tuberculosis CT imaging analysis through synthetic data augmentation via deep adversarial models | JoVE Visualize

Area of Science:

Medical Imaging
Artificial Intelligence
Computational Biology

Background:

Tuberculosis (TB) diagnosis is challenged by data scarcity and class imbalance in medical imaging datasets, particularly in resource-limited settings.
Existing datasets for TB analysis using CT scans are limited, hindering the development of robust diagnostic models.

Purpose of the Study:

To address data scarcity and class imbalance in tuberculosis CT image analysis using synthetic data augmentation.
To develop and evaluate generative adversarial networks (GANs) for synthesizing realistic TB lesion CT images.
To improve the performance of segmentation and classification models for TB detection and severity assessment.

Main Methods:

Utilized a multimodal tuberculosis dataset from Hugging Face for feature extraction from CT images and lesion regions.
Trained Deep Convolutional GAN (DCGAN) and CycleGAN models to generate synthetic CT images mimicking TB lesion morphology.
Employed U-Net architectures for segmentation and supervised networks for severity classification on GAN-augmented datasets.

Main Results:

GAN-augmented datasets demonstrated significant improvements in image fidelity and diagnostic performance metrics (FID, IS, SSIM, Accuracy, Dice, Precision, Recall).
Models trained with GAN-augmented data exhibited higher segmentation accuracy and classification precision compared to non-augmented baselines.
The study confirmed enhanced model robustness and generalizability through synthetic data augmentation.

Conclusions:

Generative adversarial network-based synthetic augmentation is a feasible and valid technique for enhancing tuberculosis lesion detection and severity assessment from CT images.
This approach is particularly valuable in environments with scarce annotated data, offering a pathway to improved diagnostic capabilities.
The findings support the integration of GANs for augmenting medical imaging datasets to overcome limitations in data availability and improve AI-driven diagnostic tools.