AI Radiology Report Translation with Doctor Review

AI Report Translation delivers doctor-verified radiology reports in the customer's language, in the cabinet, as PDF and by email.

8 report languages, more added without code

100% of translations reviewed by a radiologist before release

0 outdated translations visible after the original changes

Partner's Profile

Background

The client is a teleradiology platform in Ukraine for remote reading of CT, MRI, X-ray and mammography. Its radiologists write reports in Ukrainian on studies sent by clinics and patients. Some customers and patients read other languages: foreign clinics, patients living abroad, people asking for a second opinion. Yet the system issued every report only in its original language.

Objective

Let the customer set report languages in the order from the start and deliver a translation that a doctor has verified, in the cabinet, as PDF and by email. A machine translation of a medical document must never reach a customer unreviewed, and a translation must never outlive a change to the original report. New languages had to be added without code changes.

Solution

We added report translation into the languages specified in the order, with 8 languages at launch: English, Polish, Czech, German, Spanish, Romanian, Georgian and French. Languages live in the database, so new ones are added without code changes. Each customer organization has default languages that apply automatically to every new order, whether it comes from the web cabinet or through the API that clinics' medical information systems use. Documents uploaded by patients are auto-translated into Ukrainian for the doctor.

Translation runs as a background task through an OpenAI model with structured output against a fixed schema. The prompt is stored in the database and can be edited without a deploy. It requires the model to keep medical terminology, document structure and HTML markup; never change numbers, measurements or dates; and add, remove or interpret nothing. This is the kind of constrained LLM use we build in AI software development.

The central design decision is human-in-the-loop. The model output is only a draft. In the report editor the radiologist gets a Translations tab with the original and the translation side by side, edits the draft and publishes it. A consistency guard keeps this honest: every translation stores a hash of the source text, and if the doctor edits the original, all its translations are marked outdated and disappear from the customer's cabinet and the patient's links. About 50 dedicated automated tests cover the feature. Translations move through clear statuses:

draft: generated by the model, visible only to the doctor
published: reviewed by the doctor, visible to the customer with its own PDF; the patient gets email links that expire
outdated: the original changed, so the translation is hidden until it is reviewed again

Languages Set in the Order

The customer picks report languages when ordering, and each organization's default languages apply automatically to every new order, from the web cabinet or through the API used by clinics' medical information systems. New languages are added in the database, without code changes.

Side-by-Side Review for the Radiologist

A Translations tab in the report editor shows the original and the machine draft side by side. The radiologist edits the draft and publishes it; until then the customer and the patient see nothing.

Constrained LLM Prompt

Translation runs as a background task with structured output against a fixed schema. The prompt, stored in the database, requires keeping terminology, structure and HTML markup, never changing numbers, measurements or dates, and adding or interpreting nothing.

Source Hash Consistency Guard

Every translation stores a hash of the source text. When the doctor edits the original, all translations turn outdated and disappear from the customer's cabinet and the patient's links until they are reviewed again.