Anonymous codebase submission for SAFER Workshop in MICCAI 2026.
Standard VQA models for radiology produce short, flat answers. CORTEX instead generates a four-step chain-of-thought that mirrors the diagnostic workflow of a board-certified radiologist:
| Step | Tag | Description |
|---|---|---|
| 1 | <task> |
Restate the clinical question, modality, and available clinical context |
| 2 | <observation> |
Systematic visual assessment by anatomical region |
| 3 | <reason> |
Differential reasoning — evaluate each option or hypothesis |
| 4 | <answer> |
Final answer with rationale |
CORTEX/
│
├── prompts/
│ ├── generation/
│ │ ├── open_ended.md # Prompt for generating open-ended reasoning traces
│ │ ├── closed_ended.md # Prompt for generating closed-ended reasoning traces
│ │ └── report_gen.md # Prompt for generating radiology report reasoning traces
│ │
│ └── validation/
│ ├ ├── open_ended.md # Prompt for rubrics based scoring open-ended responses
│ ├ ├── closed_ended.md # Prompt for rubrics based scoring closed-ended responses
│ ├ └── report_gen.md # Prompt for rubrics based scoring report generation responses
│
└── dataset_samples/
│ ├── open_ended/ # Sample open-ended QA pairs with CORTEX reasoning
│ │ ├── 1.md
│ │ └── 2.md
│ ├── closed_ended/ # Sample closed-ended QA pairs with CORTEX reasoning
│ │ ├── 1.md
│ │ └── 2.md
│ └── report_gen/ # Sample report generation pairs with CORTEX reasoning
│ ├── 1.md
│ └── 2.md
│
├── expert_feedback/
│ ├── open_ended.md # Clinical experts' feedback for open-ended VQA sample in CT-RATE and CORTEX
│ ├── closed_ended.md # Clinical experts' feedback for closed-ended VQA sample in CT-RATE and CORTEX
│ └── report_gen.md # Clinical experts' feedback for report generation VQA sample in CT-RATE and CORTEX
│
├── expert_feedback/
│ └── CTRATE_samples.md # Sample CT-RATE (base dataset) QA pairs for different VQA categories in the base dataset.
Each sample file shows a side-by-side comparison of a baseline dataset response (CT-RATE) and the full CORTEX structured reasoning trace.