Beyond Imagery: AI-Enhanced Diagnostic Assistant for Cancer and Tumor Diagnosis using Radiology Imaging
This post is a technical walkthrough of our IJETS paper, Beyond Imagery: AI-Enhanced Diagnostic Assistant for Cancer and Tumor Diagnosis using Radiology Imaging (NandhaGopal S M, Rahul Sharma, Nithin M, Prajwal B R, Prashanth Kalgonda — International Journal On Engineering Technology and Sciences, Vol. XI, Issue III, March 2024). It was my final-year undergraduate research project at HKBK College of Engineering.
A radiologist looking at a CT scan is not just pattern-matching pixels against a list of diseases. She is interrogating the image — Is there a mass in the upper lobe? Is it enhancing? Has it grown since the last scan? — and answering by fusing what she sees with years of accumulated medical knowledge.
Most computer vision systems in medicine do none of that. They classify: this image is “malignant” or “benign”, with a confidence score. Useful, but rigid. The clinician cannot ask a follow-up question, and the system cannot explain itself by pointing at the literature it “learned” from.
Our paper asks a simple question: what if the diagnostic assistant could hold a conversation about the image — and ground its answers in biomedical literature at the same time? The answer we landed on is a multimodal generator (LLaVA-Med) wired to a retrieval layer over biomedical literature. Below is the full architecture, how each part works, what it scores, and where it breaks.
1. The system at a glance
The design philosophy is the same one that made parameter-efficient fine-tuning take off: keep the expensive pretrained machinery frozen, and train only the glue. The vision tower and the LLM already know how to see and speak — the adapter’s whole job is to translate image features into tokens the LLM is willing to attend to.
2. The multimodal core, piece by piece
2.1 Vision encoder — turning pixels into patch tokens
Medical images are not ImageNet photos: grayscale, high resolution, pathology in fine texture. The encoder is a pretrained vision transformer (CLIP ViT-L/14 in the LLaVA-Med lineage; ResNet-18-style CNNs are the classic baseline). It splits the scan into patches and emits one feature vector per patch:
\[\mathbf{Z}_v = f_v(I) \in \mathbb{R}^{N \times d_v}\]where $N$ is the number of patches and $d_v$ the vision width. The encoder is trained/frozen against medical imaging sets such as VQA-RAD and PathVQA, optimizing cross-entropy — so the features it surfaces are the ones that separate pathologies, not natural-image categories.
2.2 Question encoder — parsing the interrogation
The question “Is there pleural effusion, and on which side?” is tokenized and passed through a transformer encoder (RNN/LSTM in older designs). Self-attention resolves what the question is actually about — the organ (pleura), the finding (effusion), the demanded output (a side). This produces the query representation used twice downstream: fused with the image for answering, and as the retrieval query for RAG.
2.3 Multimodal fusion — the MLP adapter
This is the heart of the parameter-efficient design. Rather than co-training a giant joint model, a small two-layer MLP projects vision tokens into the LLM’s embedding space:
\[\mathbf{H}_v = \big(\mathrm{GELU}(\mathbf{Z}_v W_1 + \mathbf{b}_1)\big) W_2 + \mathbf{b}_2\]with the hidden layer deliberately smaller than the input embeddings — an information bottleneck that forces the projection to keep only what matters for answering. The first layer compresses each patch token from the vision dimension $d_v$ down to the bottleneck width; GELU is the nonlinearity that lets the mapping bend around the geometry of the language embedding space rather than staying a flat linear rotation; the second layer expands back out to the LLM’s input dimension $d_l$. After training, an LLM that has never seen an image attends to these projected tokens as if they were ordinary word embeddings — that is the whole trick.
Why an adapter and not fine-tuning the encoders themselves? Three reasons, all practical:
- Parameter budget. The adapter holds $O(d_v \cdot h + h \cdot d_l)$ weights — a few million — against tens of billions frozen across the towers. Training signal is concentrated where the modality gap actually lives.
- Catastrophic forgetting. The frozen LLM retains everything it knows about medical language and reasoning; fine-tuning it end-to-end on a few thousand QA pairs would overwrite exactly the knowledge we are borrowing.
- Feature selection on top. LLaVA-Med additionally ranks fused image/question features by cosine similarity to the query vector and keeps only the most relevant — a cheap attention-like filter that stops noisy patch tokens from diluting the question.
2.4 Answer decoder — generation under evidence
The decoder is a frozen transformer LLM (Vicuna in LLaVA-Med; the paper also discusses LLaMA-2 and GPT-3.5-class models). At each step it predicts:
\[P(y_t \mid y_{<t},\; q,\; \mathbf{H}_v,\; \mathcal{D}_q)\]Note the $\mathcal{D}_q$ — the retrieved passages. That conditioning term is the entire difference between a chatbot that remembers medicine and one that looks it up. Training minimizes cross-entropy against ground-truth answers; because only the adapter receives gradients, the LLM’s clinical language ability (and its failures) stay untouched.
3. The RAG layer — retrieval before generation
The retrieval pipeline runs in four steps:
1. Index. Biomedical literature and case studies (the PMC-15M corpus that LLaVA-Med itself was curriculum-trained on, plus curated medical databases) are chunked and embedded once, offline.
2. Query. For each incoming question, the question encoder’s representation $e_q$ is matched against document embeddings by cosine similarity:
\[\mathcal{D}_q = \operatorname{TopK}_{d \in \mathcal{C}} \; \frac{e_q \cdot e_d}{\lVert e_q \rVert \, \lVert e_d \rVert}\]3. Assemble. The prompt builder stitches together: image tokens from the adapter, the question, and the top-$k$ retrieved passages — in that order, so attention flows from evidence to question.
4. Generate. The LLM answers conditioned on the passages, which is what makes responses citable: claims in the answer trace back to retrievable sources rather than to compressed weights.
This matters double in oncology. First, for credibility — a generated claim a clinician can verify against its source is a different artifact than one she has to take on faith. Second, for freshness — treatment guidelines move faster than model retraining cycles, and retrieval absorbs that drift without touching a single weight.
4. Evaluation setup
The paper evaluates on the three standard biomedical VQA benchmarks:
| Dataset | Images | QA pairs | Character |
|---|---|---|---|
| VQA-RAD | 315 | 3,515 | Clinician-authored; head/chest/abdomen; 11 question categories; ~50% closed-ended |
| SLAKE | 642 | 7,000+ | Physician-annotated, knowledge-enhanced; segmentation masks + bounding boxes; bilingual (English subset used) |
| PathVQA | 4,998 | 32,799 | Pathology; location/shape/color/appearance; open- and closed-ended |
Metrics. Accuracy for closed-set questions; recall (fraction of ground-truth tokens present in the generated answer) for open-set questions.
Base model provenance. LLaVA-Med is curriculum-trained over the PMC-15M figure–caption corpus and fine-tuned on studies from MIMIC-CXR, reaching BLEU-4 of 0.264 and precision of 0.311 on its report-generation objective — the numbers that establish the floor our VQA task builds on.
5. Results — and the honest part
Three findings from the paper’s comparison against supervised and generative baselines:
- LLaVA-Med variants beat vanilla LLaVA across the board, with margins shifting based on language-model and vision-encoder initialization. Biomedical curriculum training transfers.
- Closed-set questions: state of the art on VQA-RAD and PathVQA. When the answer set is constrained, instruction-tuned multimodal models outperform fully supervised pipelines — they follow the task definition better.
- Open-set questions are where it gets uncomfortable. The model achieves SoTA on SLAKE but lags elsewhere. Unconstrained biomedical questions are ambiguous; a generative decoder without answer options can produce fluent, plausible-sounding answers that are simply not in the image.
Point 3 is the most useful result in the paper. Closed evaluation hides failure modes — with five options, a lucky guess and grounded reasoning score identically. Open generation exposes the gap, and it is precisely the gap retrieval is meant to close: give the decoder something concrete to condition on, and fluency stops being a liability.
6. Where this goes next
The paper’s limitation is its roadmap. Open-set performance is the real test of clinical usefulness, and the fix is not a bigger language model — it is better grounding: stronger retrieval (hybrid dense + keyword, reranking), domain-specific knowledge integration, and evaluation that rewards correct-and-supported answers over merely plausible ones. The first step of that direction is architectural: train only a lightweight projector on a frozen backbone, and let retrieval carry the domain knowledge.
Reference
NandhaGopal S M, Rahul Sharma, Nithin M, Prajwal B R, Prashanth Kalgonda, “Beyond Imagery: AI-Enhanced Diagnostic Assistant for Cancer and Tumor Diagnosis using Radiology Imaging,” International Journal On Engineering Technology and Sciences (IJETS), Vol. XI, Issue III, March 2024, pp. 41–46. ISSN (P) 2349-3968, ISSN (O) 2349-3976. Available: PDF ↗