UngLong/openm3chest-npy-v2

OpenM3Chest NPY v2

CT scan volumes (NPY float16) for LoRA fine-tuning of MedGemma 1.5-4B-IT.

Labels & tasks: UngLong/openm3chest-labels-v2

Dataset Summary

Scans	~3,793 unique CT volumes
Format	NumPy float16 (.npy)
Shape	[Z, H, W] — Z varies per scan (typically 100–300 slices)
Units	Hounsfield Units (HU)
Source	NLST via IDC

File Structure

{PatientID}/{SeriesInstanceUID}.npy

Example:

118553/1.2.840.113654.2.55.100014252814244258018667405625953775316.npy

Loading a Volume

import numpy as np
from huggingface_hub import hf_hub_download

npy_path = hf_hub_download(
    repo_id   = "UngLong/openm3chest-npy-v2",
    filename  = f"{pid}/{series_uid}.npy",
    repo_type = "dataset",
)
volume = np.load(npy_path).astype(np.float32)  # [Z, H, W], HU values

Loading with Labels

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

ds = load_dataset("UngLong/openm3chest-labels-v2", "nodule_presence")

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).astype(np.float32)  # [Z, H, W]
print(volume.shape, volume.dtype)

CT Preprocessing (MedGemma 1.5)

Per-slice RGB conversion using 3 HU windows (Yang et al. 2024):

import numpy as np
from PIL import Image

def slice_to_rgb(s: np.ndarray) -> Image.Image:
    def win(arr, lo, hi):
        return ((np.clip(arr, lo, hi) - lo) / (hi - lo) * 255).astype(np.uint8)
    r = win(s, -1225.0, 1025.0)  # bone/lung
    g = win(s,  -135.0,  215.0)  # soft tissue
    b = win(s,     0.0,   80.0)  # brain
    return Image.fromarray(np.stack([r, g, b], axis=-1), mode="RGB")

volume = np.load(npy_path).astype(np.float32)
slices = [slice_to_rgb(volume[i]) for i in range(volume.shape[0])]

Sequential Inference (Radiology)

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

Source

Built from OpenM3Chest — NLST-based chest CT dataset.
Original DICOM available via IDC (Imaging Data Commons).