VN

Research Areas

Research Interests

Detailed descriptions of my research interest areas, the motivations behind them, and the open questions I aim to address during my PhD.

Medical Image Analysis

AI-powered analysis of radiological and clinical images to support diagnosis, screening, and disease monitoring. I focus on building models that are not only accurate but also practically deployable in clinical workflows.

Why It Matters

Medical imaging is one of the highest-leverage areas for AI in healthcare. Chest X-rays, ultrasound, and MRI are routinely acquired but require expert interpretation—AI can extend access to diagnostic quality at scale, especially in low-resource settings.

Open PhD Research Questions

  • How can semi-supervised learning reduce reliance on large annotated medical image datasets?
  • What model architectures generalize best across different imaging protocols and scanner vendors?
  • How can AI-based diagnostic tools be validated for clinical deployment in low-resource settings?

Related Projects

Tuberculosis DetectionPneumonia DetectionKidney Ultrasound Classification

Explainable Artificial Intelligence (XAI)

Developing and evaluating methods that make deep learning model decisions interpretable to clinicians. This includes gradient-based attribution methods, attention visualization, and quantitative evaluation frameworks for saliency maps.

Why It Matters

Clinicians cannot trust AI systems they cannot interpret. XAI is the bridge between algorithmic accuracy and clinical adoption. Without meaningful explanations, even high-performing models remain inaccessible to the practitioners who need them.

Open PhD Research Questions

  • How can explainable AI improve trust in medical image diagnosis in clinical practice?
  • What quantitative metrics best capture the clinical relevance of model explanations?
  • How should AI explanation methods be evaluated when ground-truth clinical reasoning is unavailable?

Related Projects

Tuberculosis Detection (Grad-CAM evaluation)XAI Evaluation Study

Vision Transformers & Hybrid Architectures

Designing and analyzing Vision Transformer-based models and hybrid CNN-ViT architectures, with a focus on adapting them to the constraints of medical imaging: limited data, class imbalance, and small pathological regions.

Why It Matters

Vision Transformers offer powerful global attention mechanisms, but they require large datasets. In medical imaging, annotated data is scarce. Hybrid architectures that combine CNN efficiency with ViT expressiveness offer a practical path forward.

Open PhD Research Questions

  • What fusion strategies best integrate CNN local features with ViT global attention for medical images?
  • How can pre-training on natural images be effectively transferred to medical domains?
  • What regularization methods make ViT attention maps focus on clinically relevant regions?

Related Projects

Hybrid CNN-ViT Tuberculosis Detection

Multimodal AI for Healthcare

Exploring how combining multiple data modalities—imaging, clinical notes, lab results, patient history—can improve diagnostic AI systems beyond what single-modality models can achieve.

Why It Matters

Clinical decisions are rarely made from a single data source. Physicians integrate imaging findings with patient history, lab results, and symptoms. Multimodal AI models that reflect this integration are more realistic clinical tools.

Open PhD Research Questions

  • How can cross-modal attention effectively fuse visual and textual clinical features?
  • What architectures handle missing modalities gracefully in deployment settings?
  • How can multimodal training data be built from existing electronic health records?

Related Projects

ResearchPilot AI (multi-source integration)

Edge AI for Medical Applications

Investigating model compression, quantization, and efficient architecture design to enable AI inference on resource-constrained devices—critical for point-of-care diagnostics in low-resource clinical settings.

Why It Matters

High-end GPUs are not available in most hospitals in developing countries. If medical AI is to have meaningful impact globally, models must run efficiently on affordable hardware, including mobile devices and embedded systems.

Open PhD Research Questions

  • What quantization strategies preserve diagnostic accuracy for medical image classifiers?
  • How can knowledge distillation create compact models that retain clinically meaningful features?
  • What are the trade-offs between model compression and explanation quality in XAI methods?

Related Projects

YOLOv8 Emergency Detection API (real-time deployment)

AI-assisted Research Tools

Building intelligent tools that support the research process itself—literature discovery, knowledge synthesis, experimental tracking, and scientific writing assistance.

Why It Matters

Research productivity is bottlenecked by the manual effort of finding, reading, and synthesizing literature. AI tools can significantly reduce this overhead, allowing researchers to focus on ideation and experimentation.

Open PhD Research Questions

  • How can large language models support structured literature reviews without hallucinating citations?
  • What interfaces best support researcher-AI collaboration in hypothesis generation?

Related Projects

ResearchPilot AI Platform