Inspiration
Laziness. Laziness is what inspired me to build what I have built and I am sure it will be helpful not just for me, but for other Radiologists too.
What it does
AMMAN stands for Augmented Medical Mismatch Analysis & Notification.
AMMAN is an independent post-report verification layer for ConchRad, my existing radiology reporting application. ConchRad generates Radiology reports in 20-30seconds based on image annotations and measurements in personalized templates. So, AMMAN feature checks:
- ConchRad's structured extracted evidence, such as measurements, gestational age, presentation, placenta, and anatomy notes. This is an example for a Obstetric report(Pregnancy scan)
- The generated draft report
Using GPT-5.6 in strict mode, AMMAN notifies the Radiologist about potentially unsupported statements, evidence-versus-draft mismatches, uncertainty differences, and important omitted evidence. It produces review flags only; it does not rewrite, approve, sign, or silently alter the report. The radiologist remains the final decision-maker.
How I built it
ConchRad was alreadyt built. AMMAnb feature was an added second layer correction feature built using CODEX and GPT 5.6 The desktop client sends the structured extraction and report draft to a private Flask service after report generation. The service uses the OpenAI API and GPT-5.6 to return structured review findings. I used Codex to design, implement, test, and safety-harden the integration, including timeout handling, strict review behavior, and safeguards against OCR shortcut codes alone creating fibroid findings.
The demonstration uses sample, de-identified test data only.
Challenges I ran into
I am a Doctor by profession with zero knowledge about Coding/programming. So there was no challenge. I enjoyed the whole process. :)
What I learned
This is awesome and I am inspired to build more. Medical/Radiology related softwares.
Log in or sign up for Devpost to join the conversation.