UngLong/radiology-ready-v2

Radiology Ready v2

Training-ready dataset for fine-tuning the Radiology agent of an AI Medical Department, built on MedGemma 1.5-4B-IT.

Each row is a (prompt, response) pair ready for SFT. CT scan images are not stored here — they live in UngLong/openm3chest-npy-v2 and are fetched at training time via {pids}/{keys}.npy.

Dataset Summary

Rows	500
Screening rows	302 (Pool A + Pool B)
Detail rows	198 (Pool A — nodule present)
Unique CT volumes	477 (Pool A scans appear in both screening & detail)
Output format	JSON (multi-task)
Source	NLST via IDC

Two-Step Radiology Design

The Radiology agent uses sequential inference, mirrored in the training data via task_type:

Step 1 — Screening  (task_type="screening", all scans):
    chest_abn_54–61 + nodule_presence  → 8 findings

Step 2 — Detail     (task_type="detail", only if nodule_presence = Yes):
    nodule_location / attenuation / margin / size  → 4 characteristics

A single LoRA adapter is trained on both row types (distinguished by the prompt). At inference: run screening first; if nodule_presence = Yes, run the detail prompt on the same volume.

Schema

Field	Type	Description
keys	string	SeriesInstanceUID — maps to {pids}/{keys}.npy in the NPY repo
pids	string	Patient ID
prompt	string	Clinical record + questions (with allowed values) + JSON output instruction
response	string	JSON object mapping each task → text answer
task_type	string	"screening" or "detail"

Example row (screening)

{
  "keys": "1.2.840.113654...",
  "pids": "118553",
  "task_type": "screening",
  "response": "{\"chest_abn_54\": \"No\", ..., \"nodule_presence\": \"Yes\"}"
}

Tasks & Labels

Screening (8 tasks, binary No/Yes)

Task	Finding
chest_abn_54	Atelectasis (segmental or greater)
chest_abn_55	Pleural thickening or effusion
chest_abn_56	Non-calcified mass/adenopathy ≥10 mm (hilar/mediastinal)
chest_abn_57	Chest wall abnormality (bone destruction, metastasis)
chest_abn_58	Consolidation
chest_abn_59	Emphysema
chest_abn_61	Fibrosis / honeycombing / reticulonodular opacities
nodule_presence	Lung nodule present

Detail (4 tasks, multi-class)

Task	Classes
nodule_location	Right Upper / Right Middle / Right Lower / Left Upper / Left Lower Lobe
nodule_attenuation	Solid / Ground Glass / Others
nodule_margin	Spiculated (Stellate) / Smooth / Poorly defined / Unable to determine
nodule_size	≤4mm / 4–6mm / 6–8mm / 8–15mm / 15–30mm / >30mm

Label Balance (screening, positive rate)

This subset re-selects positive cases from the NPY-ready pool to mitigate the severe class imbalance of NLST screening findings (rare abnormalities):

Task	Positive %
chest_abn_55	54%
nodule_presence	71%
chest_abn_59	48%
chest_abn_61	37%
chest_abn_56	26%
chest_abn_54	9%
chest_abn_58	6%
chest_abn_57	2% ⚠️

Limitation: chest_abn_57 (chest wall abnormality) has ~0.1% prevalence in NLST; only a handful of positive cases exist even in the full cohort. Evaluate it with caution. Use F1 / recall per task rather than accuracy, since several findings remain imbalanced.

Usage

import json
import numpy as np
from datasets import load_dataset
from huggingface_hub import hf_hub_download

ds = load_dataset("UngLong/radiology-ready-v2", split="train")
row = ds[0]

# Fetch the CT volume for this row
npy_path = hf_hub_download(
    repo_id="UngLong/openm3chest-npy-v2",
    filename=f"{row['pids']}/{row['keys']}.npy",
    repo_type="dataset",
)
volume = np.load(npy_path).astype(np.float32)  # [Z, H, W], HU values

print(row["task_type"])          # "screening" or "detail"
print(row["prompt"])
print(json.loads(row["response"]))   # {task: answer, ...}

CT Preprocessing (MedGemma 1.5)

Per-slice RGB via 3 HU windows (MedGemma 1.5 technical report):

def hu_window(s, lo, hi):
    return ((np.clip(s, lo, hi) - lo) / (hi - lo) * 255).astype(np.uint8)

# R: (-1024, 1024)  bone/lung   G: (-135, 215)  soft tissue   B: (0, 80)  brain/vascular

Sample up to 85 axial slices equidistantly; the processor resizes to 896×896.

Source

Built from OpenM3Chest — NLST-based chest CT dataset with clinical labels. CT images available via IDC.