Van-Ninh Ha

Van Ninh Ha

Manuscript in Preparation

This manuscript systematically evaluates explainability methods—including Grad-CAM, Grad-CAM++, and SHAP—applied to deep learning classifiers for medical image diagnosis. We define quantitative evaluation metrics for saliency map quality and assess alignment between model explanations and radiologist annotations on chest X-ray datasets.

Van Ninh Ha, Co-author A, Co-author B

International Conference on [Venue — To be confirmed]

We propose a hybrid architecture combining convolutional neural networks and Vision Transformers for tuberculosis detection in chest X-ray images. The model leverages local feature extraction from CNNs and global context modeling from ViTs to improve detection accuracy, particularly in cases with subtle pathological signs. Experimental results on publicly available datasets demonstrate competitive performance compared to existing state-of-the-art methods.

Van Ninh Ha, Co-author A

Workshop on [Venue — To be confirmed]

This work investigates the effect of various data augmentation strategies on deep learning models for pneumonia detection from chest X-ray images. We systematically evaluate augmentation techniques including random cropping, rotation, elastic deformation, and CutMix across multiple CNN architectures. Our results show that targeted augmentation pipelines significantly improve generalization under limited training data conditions.