// Anthropic API · One multimodal call

DICOM Study Analysis

This demo runs a two-step offline pipeline: first, it preprocesses anonymized DICOM series and sends 10 slices per series in one direct Anthropic multimodal request; second, it runs a per-series summarization pass to add concise narrative takeaways. The webpage then renders static JSON with start/middle/end thumbnails per series. No runtime API calls occur in this page.

Study ID—

Series—

Selected slices—

Generated—

Study-level interpretation

Possible findings

Uncertainties

Per-series analysis

// Technical summary

What this demo uses

pydicom Anthropic API Static JSON JavaScript

The browser loads generated JSON and PNG files only. DICOM parsing, slice extraction, and model prompting happen offline in Python.

Methodology: Step 1 runs one multimodal call using 10 evenly spaced slices per series plus curated DICOM metadata; Step 2 runs one concise-summary call per series and appends `concise_summary` text to each series result.
Confidence policy: Findings are listed with a model-reported confidence score.
Toolsets/technology used: Python scripts (`generate_dicom_study_demo.py`, `generate_dicom_series_summaries.py`), pydicom, Pillow, NumPy, Anthropic API, and vanilla JavaScript.