Medical Report Analysis using RAG

About the medical report analysis using RAG M.Tech project

A blood test report is full of short forms like HbA1c, SGPT and TSH, with numbers and ranges most patients cannot read. This medical report analysis using RAG M.Tech project helps them understand it. The patient photographs the report in an Android app, OCR reads each test, and a retrieval augmented generation pipeline explains every value in simple English or Hindi, with a numbered source behind each sentence.

The research question is about trust: how far do better retrieval and a sentence-level source check cut down statements that no trusted source supports?

Research behind it

Base paper. The study starts from Lab-AI: Using Retrieval Augmentation to Enhance Language Models for Personalized Lab Test Interpretation in Clinical Medicine, presented at IEEE ICHI 2025. It finds which personal factors such as age and sex change a test's normal range, retrieves the matching MedlinePlus passage from a vector database, and shows that retrieval makes language models far more accurate on range questions.

Research gap. Lab-AI is tested on typed question sets. A real patient starts from a crumpled photo with mixed units and printed ranges. Retrieval in the base work is dense only, which can miss exact test names such as "SGPT". And a patient needs to see where each sentence came from, in a language they read comfortably.

Proposed system and novelty. OCR and a parser pull out test, value, unit and printed range, and units are normalised. Ranges are picked for the patient's age and sex. BM25 and dense MiniLM search run together, and a cross-encoder reorders the passages. The language model must cite a passage number for every sentence, and an NLI model removes any sentence that its cited passage does not support. Output is in English or Hindi, and an offline template mode works without internet in the viva hall.

Models compared. A plain language model without retrieval, the base paper's dense pipeline re-implemented, and the offline template mode, against the proposed pipeline. Metrics include field extraction F1, range correctness, recall at k, faithfulness and unsupported statements per report. The ablation removes the reranker, removes the NLI filter and switches to dense-only search.

The same retrieval idea over college documents is used in the RAG chatbot project, if you want a non-medical version.

How it works

• The patient photographs the report, and the image is straightened and cleaned.
• Tesseract reads the text, and rules extract test names, values, units and ranges.
• The patient can fix any wrong value on screen.
• Each value is marked low, normal or high for the patient's age and sex.
• Hybrid retrieval fetches passages; the model writes a cited explanation.
• The NLI check removes unsupported sentences before the answer is shown.
• The patient taps a citation to read the source, or shares the report with a doctor.

Users and modules

There are three roles. The patient scans reports, reads explanations, asks questions, tracks values over time and downloads a PDF. The doctor sees reports that patients share and adds a short note. The admin manages users and records, edits knowledge base passages, checks the comparison page and stores the API key and language settings. The Android side is Java, and the server is Flask with MySQL.

What you get

• Phase 1 Research: Base paper study: Lab-AI retrieval augmentation (IEEE ICHI, 2025); Literature survey of 15-20 recent RAG and medical NLP papers; Research gap, novelty and objectives; MedlinePlus knowledge base and labelled report set; Synopsis and Phase-1 report.
• Phase 2 Implementation: Android app (Java) with camera and chat; Flask API, doctor and admin panels; OCR, parsing and unit normalisation pipeline; Hybrid retrieval, reranker and cited generation; NLI faithfulness filter and offline mode; MySQL database and knowledge base files.
• Results: OCR field extraction table; Recall at k for dense and hybrid retrieval; Faithfulness and unsupported statements per pipeline; results.xlsx and all graphs in high resolution.
• Thesis and paper: Thesis written for you (about 90 pages); IEEE conference paper draft; Journal paper draft; Publication guide.
• Reviews, viva and support: PPTs for Review 1, Review 2 and final viva; Viva book with 50+ questions; A1 poster and project diary; Demo and code walkthrough videos; 1-on-1 setup, 2 revisions and support till viva.

Who can use this project

M.Tech and M.E. students of CSE and IT who want a generative AI dissertation with a measurable research question will find this useful. It suits searches for a hybrid retrieval BM25 dense reranking project or an LLM hallucination evaluation project, and the Android app gives your guide something practical to test.