Radiology Reporting Model — Imaging Group, Queensland, Australia
The Situation
A radiology group with 4 imaging centres across Queensland had been trialling a general-purpose LLM for report drafting for 6 months before approaching Amorisoft. The trial had not gone well.
The model used imprecise clinical language. It could not follow the group's structured reporting templates, which varied by modality across CT, MRI, and plain film. It got Australian spelling conventions wrong often enough that radiologists had stopped trusting the drafts and were writing from scratch anyway. The tool was still running, but nobody was using it.
What the group did have was 22,000 de-identified historical reports filed over 9 years, across all 4 centres and all modalities. That was the starting point.
What Amorisoft Did
Amorisoft ran a 4-week data preparation process to clean, de-identify, and structure the report corpus into a training dataset. Each report was tagged by modality, centre, radiologist seniority, and report type. Structured reporting templates were encoded as format constraints rather than free-text examples, so the model would learn to follow them rather than approximate them.
Fine-tuning ran on a base model selected for performance on clinical long-form text. The de-identification pipeline was reviewed by the group's privacy officer before any data left the group's own storage environment. Nothing moved to external infrastructure until sign-off was in place.
RLHF ran across 4 feedback cycles. Three senior radiologists from different centres acted as raters, which mattered: reporting style varied enough between centres that single-rater feedback would have biased the model toward one site's conventions. By cycle 2 draft acceptance was above 80%. Cycle 4 brought it to 91%.
The same model serves all 4 centres. Centre-specific template variants are handled through prompt configuration rather than separate fine-tuning runs, which kept the maintenance overhead low after deployment.
Implementation Note
The oldest reports in the corpus, covering years 1 through 3, used a different template structure from the current standard. Including them in training without flagging the difference caused early model outputs to mix old and new formatting conventions. Amorisoft's data team caught this in evaluation week 1 and rebuilt the dataset with era tags. It added 5 days to the preparation phase.
Results
Report drafting time dropped by 64% across all 4 centres within 8 weeks of go-live. Radiologists accepted 91% of drafts without revision by the end of the RLHF process, a figure that held in production. The de-identification pipeline processed all 22,000 training reports with zero patient data exposure confirmed through two independent audits. All 4 centres run on a single model with no per-centre retraining required.