Research

Two DOI-archived papers, with a real control group

The featured study tests a real question with a proper baseline and significance testing, not one flattering number.

Featured research
2026· Independent research · DOI-archived on Zenodo

MedInsight: Evaluating Retrieval-Augmented Vision-Language Models for Evidence-Grounded Medical Image Understanding

Does giving a medical vision-language model retrieved similar-case evidence actually make it more accurate — or does it just look like it should?

Abstract

MedInsight tests whether retrieval-augmented prompting improves a vision-language model's accuracy on medical visual question answering, using only openly licensed, non-credentialed data. A CLIP-embedded FAISS index over ~90,000 ROCOv2 radiology image-caption pairs supplies retrieved evidence captions to a pretrained BLIP-2 model, and that retrieval-augmented condition is compared against the same model answering with no retrieved evidence at all — same checkpoint, same decoding settings, same scoring code, on the official VQA-RAD test split (451 clinician-authored questions), with paired bootstrap significance testing rather than a single reported delta.

Contributions

  • A fully reproducible retrieval-augmented medical VQA pipeline built entirely on open, non-credentialed datasets — no data-use agreement required to run it end to end
  • A controlled baseline-vs-RAG comparison that shares one generation code path between both conditions, so any measured difference can be attributed to retrieval itself rather than an implementation quirk
  • Statistical rigor beyond a raw accuracy delta: paired bootstrap significance testing (10,000 resamples) plus a qualitative improved/regressed/unchanged error breakdown

Methodology

01

CLIP-embedded FAISS flat inner-product index over the full ROCOv2 corpus (~90K radiology image-caption pairs)

02

Shared generation code path (BaselineVLM.generate_from_prompt) for both the baseline and retrieval-augmented BLIP-2 conditions — identical tokenization, decoding, and scoring

03

Evaluation on VQA-RAD's official 451-example held-out test split, scored by exact match (closed questions) and BLEU-4/ROUGE-L (open questions)

04

Paired bootstrap significance testing (10,000 resamples) plus a sweep over retrieval depth k ∈ {1, 3, 5, 10}

VQA-RAD test exampleimage + clinical questionBaseline pathno retrieval usedQuestion-only promptno evidence blockRetrieval-augmentedCLIP image encoderFAISS index searchROCOv2 · ~60K pairs · cosine simPrompt + evidencetop-k captions prependedShared generation — BLIP-2greedy decode · identical for both conditionsAnswer + scoring · exact match · BLEU-4 · ROUGE-L

Key results

Closed-question accuracy

43.0%47.8% (+4.8 pts)

Overall exact match

26.8%29.5% (+2.7 pts)

Significance (closed EM)

p = 0.1848 — not significant at n = 451

Best retrieval depth

k = 1 (accuracy declines as k increases)

Datasets

  • ROCOv2 (retrieval corpus)
  • VQA-RAD (evaluation benchmark)

Stack

Python
PyTorch
BLIP-2
CLIP
FAISS
Retrieval-Augmented Generation
Vision-Language Models
Hugging Face Transformers

Findings

  • Retrieval augmentation improved every reported metric at the default depth (k = 5), but paired bootstrap testing found neither the closed-question nor open-question improvement statistically significant — an honest result more useful than an overstated one
  • Accuracy peaked at the smallest retrieval depth (k = 1) and declined monotonically through k = 10, suggesting excess retrieved evidence can dilute rather than reinforce a base model's answer
  • Error analysis showed both conditions failed on the majority of questions (58%) — the underlying VLM's zero-shot capability on radiology imagery, not retrieval, is the primary bottleneck

Limitations, stated plainly

  • VQA-RAD's 451-example test split is small, which is why significance testing was treated as a requirement rather than an optional check
  • The retriever matches only on image similarity — the question text never participates in retrieval — a plausible confound with the k-depth findings
  • A single BLIP-2 checkpoint was evaluated at a single scale; no clinician review of answer quality was performed, so results describe this pipeline, not a validated clinical tool

Earlier work

Other published research

2024 – 2025
Independent research · DOI-archived on Zenodo

AI-Based Depression Detection from Social Media Using LSTM Networks

This research investigates whether Long Short-Term Memory (LSTM) networks can pick up linguistic and behavioral indicators of depression in social media text more reliably than simple bag-of-words baselines. NLP preprocessing (tokenization, cleaning, stopword removal, embedding generation) feeds a sequence model trained to classify posts as depression-indicative or not, with class-imbalance handling and standard classification metrics.

Python
TensorFlow/PyTorch
LSTM
NLP
Pandas
NumPy