Does your pathologist refuse to use a computer? No problem! You’ll never have to enter paper documents manually again. And yet you still have control over the accuracy of the data.

ProstaPaper: from paper-based pathology reports to a digital registry

What the project does

ProstaPaper was created during the 24-hour Rakathon 2026 at the Masaryk Memorial Cancer Institute in Brno. It helps convert Czech pathology reports following prostate removal and manually annotated anatomical diagrams into structured records for the COMPARE register in REDCap. The aim is to reduce manual transcription and prepare data for research — not to make a diagnosis or recommend treatment.

It processes the text and drawings separately, then proposes how they should be linked. It preserves source evidence, missing data and discrepancies. The result must be checked by a doctor before export.

What data did we use

We worked with 80 anonymised hackathon scans: 40 reports and 40 corresponding drawings based on the PI-RADS v2.1 template. The reports contain, for example, lesion dimensions, Gleason score, Grade Group, margin status and pT/L/V/R classification. The drawings show the location of findings in individual sections, handwritten notes and dimensions.

We also utilised a map of 41 sectors, code lists and REDCap import files. The automated tests also use synthetic reports and model answers. We did not train our own model; extraction utilises Gemini and local image processing.

What we managed to achieve

• Preparation of scans: local OCR reads the hackathon ID, links the report to the drawing, assigns a stable random identifier and masks the original ID in copies. The documented run matched 40 out of 40 cases with no detected mix-ups. However, this is not a universal anonymisation method.
• Message extraction: Gemini first transcribes the text and then generates structured findings. Values are checked against citations; ambiguities remain for review.
• Image processing: Template alignment, separation of handwritten strokes, repeated reading by Gemini, comparison with OpenCV, and proposals for sectors and groups of characters are implemented.
• Merging results: the central module pairs findings with the drawing, retains conflicts and can incorporate the doctor’s recorded decisions.

The process of extracting information from the report into a joint proposal is documented across three cases. This is an integration trial, not a measurement of medical accuracy. The entire workflow, up to medical approval and verified import, has not been completed. The original review interface was removed due to a redesign.

What should follow

Complete the interface with scans, a clickable map and a queue of ambiguities; link actual analysis results and verify the CSV import into the test REDCap instance. Prepare 5–8 assessment sheets for the images and obtain human evaluations. Subsequently, measure accuracy against results confirmed by doctors.

Before going live, the following issues remain to be resolved: secure patient linkage, protection of identifiers, and hospital-approved deployment, including conditions for external processing.

Summary based on the code and documentation for the main branch as at 17 October 2026; I have not run any tests here.