TorchXRayVision

Many datasets (like NIH_Dataset and CheXpert_Dataset) support pixel-level segmentation masks. When initializing the dataset, set pathology_masks=True.

Each sample returned by the dataset will then include a "pathology_masks" key, which maps pathology names to binary pixel arrays of the same spatial size as the image ("img").

import torchxrayvision as xrv

ds = xrv.datasets.NIH_Dataset(
    imgpath="/data/NIH",
    pathology_masks=True,
)
sample = ds[0]
print(sample.keys())          # dict_keys(['img', 'lab', 'pathology_masks', ...])
print(sample["pathology_masks"].keys())  # e.g. {'Atelectasis': array(...)}

If a model prediction for a specific pathology is consistently around 0.5, it likely indicates that the specific label was not included in the training set for that weight configuration (an untrained head).

Only interpret labels that are explicitly present in model.pathologies. You can inspect the raw logits for all pathologies using the following pattern:

import torchxrayvision as xrv
import torch

model = xrv.models.DenseNet(weights="densenet121-res224-all")
model.eval()

img = xrv.utils.load_image("path/to/xray.jpg")
img = xrv.datasets.XRayCenterCrop()(img)
img = xrv.datasets.XRayResizer(224)(img)
img_tensor = torch.from_numpy(img).unsqueeze(0)  # [1, 1, 224, 224]

with torch.no_grad():
    output = model(img_tensor)[0].cpu().numpy()  # shape: (num_pathologies,)

for pathology, score in zip(model.pathologies, output):
    print(f"{pathology}: {score:.4f}")

TorchXRayVision provides a unified interface for various chest X-ray datasets through a common base class xrv.datasets.Dataset. Every dataset class exposes three core attributes that allow you to access labels and metadata consistently:

• pathologies: An ordered list of label names (strings).
• labels: A 2-D NumPy array of shape (samples, pathologies) containing values 1, 0, or NaN.
• csv: A Pandas DataFrame containing per-image metadata.

To load a dataset, you typically only need to provide the path to the image directory.

import torchxrayvision as xrv

d = xrv.datasets.NIH_Dataset(imgpath="/path/to/images")

TorchXRayVision provides specialized transform objects designed for X-ray images. These transforms operate on (1, H, W) float32 NumPy arrays (the standard output format for xrv.datasets.Dataset.__getitem__ calls) and are compatible with torchvision.transforms.Compose.

from torchvision import transforms
import torchxrayvision as xrv

# Example of composing X-ray specific transforms
transform = transforms.Compose([
    xrv.datasets.XRayResizer(224),
    xrv.datasets.XRayCenterCrop(224),
])

All models in torchxrayvision follow a standard interface defined in xrv.models.Model. The primary method for inference is forward(x), where x is the input tensor. Models are typically initialized with specific weights and optional thresholding parameters.

# Conceptual usage of the Model interface
import torch
from xrv.models import Model

# Assuming a model instance 'model' is initialized
# x should be a tensor of shape (batch_size, channels, height, width)
x = torch.randn(1, 1, 224, 224)
output = model.forward(x)

Each dataset object provides access to pathologies, labels, and metadata via a pandas DataFrame. When using xrv.datasets.Subset_Dataset or xrv.datasets.Merge_Dataset, these fields are preserved.

Core Fields

• .pathologies: A list of pathology names contained in the dataset.
• .labels: A tensor/array containing 1, 0, or NaN for each label in .pathologies.
• .csv: A pandas DataFrame containing the metadata.

Common CSV Metadata Fields

• csv.patientid: Unique identifier for samples.
• csv.offset_day_int: Integer time offset in days.
• csv.age_years: Patient age in years.
• csv.sex_male: Boolean/indicator if patient is male.
• csv.sex_female: Boolean/indicator if patient is female.

To ensure deterministic behavior when creating datasets and models, set the seeds for random, numpy, and torch (including CUDA if applicable) before initialization.

import torch, numpy as np, random
seed = 0
random.seed(seed)
np.random.seed(seed)
torch.manual_seed(seed)
if torch.cuda.is_available():
    torch.cuda.manual_seed_all(seed)

You can install the torchxrayvision package using the standard Python package manager.

pip install torchxrayvision