UngLong/openm3chest-labels-v2

OpenM3Chest Labels v2

Stratified label subset of OpenM3Chest for LoRA fine-tuning of MedGemma 1.5-4B-IT.

CT scan images (NPY): UngLong/openm3chest-npy-v2

Dataset Summary

Agent	Tasks	Train samples	Test samples
Radiology	chest_abn_54–61, nodule_presence/location/attenuation/margin/size	~350/task	full
Cardiology	CVD_diagnosis, CVD_mortality	1,500/task	3,759/task
Oncology	lung_cancer_risk	2,000	10,308

Training subset: stratified sampling — binary tasks ~50/50, all categorical classes covered.
Test split: full original test set (unfiltered) for proper evaluation.

Using the Dataset

from datasets import load_dataset

# Radiology — chest abnormality detection
ds = load_dataset("UngLong/openm3chest-labels-v2", "chest_abn_54")
print(ds["train"][0])

# Cardiology — CVD diagnosis
ds = load_dataset("UngLong/openm3chest-labels-v2", "CVD_diagnosis")

# Oncology — lung cancer risk
ds = load_dataset("UngLong/openm3chest-labels-v2", "lung_cancer_risk")

Schema

Each record contains:

Field	Type	Description
keys	string	SeriesInstanceUID — maps to {pids}/{keys}.npy in NPY repo
pids	string	Patient ID
clinical_data	string	JSON-serialized clinical information
questions	list[string]	10 candidate questions for this task
labels	int64 / string	Task label (see below)
answer_dict	string	JSON-serialized dict mapping label index → human-readable answer text

Tasks & Labels

Radiology

Config	Label type	Classes
chest_abn_54	int (0/1)	0 = No, 1 = Yes
chest_abn_55	int (0/1)	0 = No, 1 = Yes
chest_abn_56	int (0/1)	0 = No, 1 = Yes
chest_abn_57	int (0/1)	0 = No, 1 = Yes
chest_abn_58	int (0/1)	0 = No, 1 = Yes
chest_abn_59	int (0/1)	0 = No, 1 = Yes
chest_abn_61	int (0/1)	0 = No, 1 = Yes
nodule_presence	int (0/1)	0 = No nodule, 1 = Nodule present
nodule_location	int (0–4)	0=RUL, 1=RML, 2=RLL, 3=LUL, 4=LLL
nodule_attenuation	int (0–2)	0=Solid, 1=Ground Glass, 2=Others
nodule_margin	int (0–3)	0=Spiculated (Stellate), 1=Smooth, 2=Poorly defined, 3=Unable to determine
nodule_size	int (0–5)	0=≤4mm, 1=4–6mm, 2=6–8mm, 3=8–15mm, 4=15–30mm, 5=>30mm

Cardiology

Config	Label type	Classes
CVD_diagnosis	int (0/1)	0 = No, 1 = Yes
CVD_mortality	int (0/1)	0 = Low risk, 1 = High risk

Oncology

Config	Label type	Description
lung_cancer_risk	string (JSON)	Dict with y (bool), y_seq (6-year sequence), y_mask

import json
record = ds["train"][0]
label = json.loads(record["labels"])
print(label["y"])      # True/False — cancer within follow-up
print(label["y_seq"])  # [0,0,0,1,0,0] — year-by-year occurrence

Loading NPY Files

import json
import numpy as np
from huggingface_hub import hf_hub_download

record = ds["train"][0]
npy_path = hf_hub_download(
    repo_id   = "UngLong/openm3chest-npy-v2",
    filename  = f"{record['pids']}/{record['keys']}.npy",
    repo_type = "dataset",
)
volume = np.load(npy_path)  # shape: [Z, H, W], dtype: float16

Sequential Inference (Radiology)

The Radiology agent uses sequential inference:

Step 1 — Screening:  chest_abn_54–61 + nodule_presence
Step 2 — If nodule_presence = Yes:
          nodule_location, nodule_attenuation, nodule_margin, nodule_size

Training data reflects this:

• Pool A (198 scans): all 12 tasks, nodule_presence = 1
• Pool B (150 scans): screening tasks only, nodule_presence = 0

Source

Built from OpenM3Chest — NLST-based chest CT dataset with clinical labels.
CT images available via IDC (Imaging Data Commons).