Abdul Manaf

My MS thesis introduced an ROI-guided multimodal pipeline for generating lesion descriptions from medical images. It benchmarked MedFlamingo, MedLLaVA, MedGemma, and Qwen2-VL under full fine-tuning, LoRA, and zero-shot settings. The study found architecture-dependent effects of region guidance and a consistent gap between semantic similarity and factual correctness, which standard metrics often hide.

Earlier work spans fracture detection in pediatric radiographs, self-supervised transformers for chest X-rays, explainable aerial-image analysis for flood response, and cross-lingual evaluation of LLMs.

• Vision-language models for medicine

  Region-grounded report and lesion description generation; parameter-efficient adaptation of medical VLMs.

• Trustworthy & explainable AI

  Measuring factual correctness beyond surface similarity; saliency and LIME-based explanations.

• RAG & agentic LLM systems

  Domain-grounded retrieval pipelines, multi-agent workflows, and evaluation of faithfulness and relevance.

• Detection in clinical imaging

  YOLO-family detectors and CNN/transformer classifiers for radiograph abnormality detection.