Body charts from CT across adulthood

This repository accompanies our work on whole-body CT reference charts (organ volume and tissue attenuation in Hounsfield Units) across adulthood, built from routine clinical CT using automated multi-organ segmentation and evidence-grounded report filtering.

Interactive reference charts

• https://ai-med.de/body-charts-v2/

Evidence-grounded, cross-model LLM filtering

LLM report filtering (batch scripts)

The folder src/ includes lightweight bash wrappers and Python entry points that were used to run the evidence-grounded, cross-model LLM filtering at scale.

• run_batch_extract.sh
  Runs the extraction stage (structured JSON summaries) with vLLM using guided decoding. Model selection is controlled via --model (e.g., medgemma), and prompts are provided via --instruction-file (e.g., prompts/instruct_extract_en.txt).

• run_batch_verify.sh
  Runs the verification stage on disputed cases (e.g., disputed_summaries/) and produces a JSONL file of verified decisions. This corresponds to the second-stage cross-checking of extracted findings using a separate verification prompt (e.g., prompts/instruct_verify_en.txt).

• batch_process_reports_json2_en.py
  Python driver for the extraction stage.

• batch_process_verify.py
  Python driver for the verification stage.

Core LLM inference & model tooling

• vllm==0.8.5.post1
• torch==2.6.0
• torchvision==0.21.0+cu126
• torchaudio==2.6.0
• transformers==4.52.4
• accelerate==1.5.2
• safetensors==0.5.3
• tokenizers==0.21.1
• huggingface-hub==0.32.3
• xgrammar==0.1.18

Quantization / memory-efficient inference (optional, depending on models)

• bitsandbytes==0.45.3
• xformers==0.0.29.post2
• flash-attn==2.7.3
• cupy-cuda12x==13.4.1

GAMLSS model fitting (minimal reference implementation)

The folder gamlss/ contains two GAMLSS model-fitting routines used in this work. The code is intended to document the core modeling algorithm (distributional regression with candidate fractional-polynomial search and information-criterion selection).

Included scripts

• fit_gamlss_model_volume.R
  Fits organ volume reference models using GAMLSS (default family: GG).
  Candidate models are defined by a small grid of fractional-polynomial (FP) orders for mu and sigma (with nu held constant). The best candidate is selected by AIC and BIC, and “final” models are refit using the implied FP powers expressed via bfp(Age, powers=...). The model adjusts for Sex, Manufacturer, Contrast and includes a random Study effect (optional kvp).

• fit_gamlss_model_ST1.R
  Fits attenuation (HU) models using the ST1 family.
  As above, the routine evaluates a small set of FP candidates, selects the BIC-best specification, and refits a final model using bfp() powers. The model adjusts for Sex, Manufacturer, includes a random Study effect, and can optionally include kvp.