PatchChestCT

Reference implementation for the PatchChestCT dataset — a large-scale 3D patch-level annotation dataset for localizing multiple abnormalities in chest CT.

Dataset:

---

Dataset Overview

PatchChestCT provides patch-level binary labels for 9 clinically significant abnormalities across 2,201 chest CT volumes derived from the CT-RATE dataset.

Each annotation is stored as a binary NumPy array (.npz) with shape (24, 12, 12), corresponding to the 3D patch grid of the CT volume after standard preprocessing.

Annotated abnormalities:

Label file	Abnormality
arterial_wall_calcification.npz	Arterial wall calcification
coronary_wall_calcification.npz	Coronary artery calcification
pericardial_effusion.npz	Pericardial effusion
hiatal_hernia.npz	Hiatal hernia
lymphadenopathy.npz	Lymphadenopathy
atelectasis.npz	Atelectasis
lung_opacity.npz	Lung opacity
consolidation.npz	Consolidation
bronchiectasis.npz	Bronchiectasis

---

Obtaining the Data

Step 1 — Get CT-RATE source volumes

The source CT volumes are not included in the PatchChestCT Zenodo deposit. You must obtain them independently from the official CT-RATE repository:

1. Go to: https://huggingface.co/datasets/ibrahimhamamci/CT-RATE
2. Request gated access and agree to the Data Usage Agreement (DUA).
3. Download and preprocess the volumes following the CT-RATE documentation.

Step 2 — Download PatchChestCT annotations

Download annotations-train.zip and annotations-valid.zip from Zenodo:

https://doi.org/10.5281/zenodo.16812648

Step 3 — Align annotations with source volumes

Each subdirectory in the annotation zips is named using the original CT-RATE volume identifier (e.g., train_1003_a_1/), which directly corresponds to the CT-RATE naming convention.

import numpy as np
import os

def load_annotation(annotation_dir, volume_id, abnormality):
    """Load a patch-level annotation for a given volume and abnormality."""
    path = os.path.join(annotation_dir, volume_id, f"{abnormality}.npz")
    if not os.path.exists(path):
        return None  # abnormality not present in this volume
    return np.load(path)["arr_0"]  # shape: (24, 12, 12), dtype: bool

# Example
ann = load_annotation("./annotations-train", "train_1003_a_1", "arterial_wall_calcification")
print(ann.shape)   # (24, 12, 12)
print(ann.sum())   # number of positive patches

---

Baselines

Data preparation

project_root/
├── train_preprocessed/     # CT-RATE training volumes (obtain from CT-RATE)
├── valid_preprocessed/     # CT-RATE validation volumes
├── annotations-train/      # PatchChestCT train annotations (from Zenodo)
├── annotations-valid/      # PatchChestCT valid annotations (from Zenodo)
├── train_reports.csv
├── dataset_radiology_text_reports_validation_reports.csv
├── dataset_multi_abnormality_labels_train_predicted_labels.csv
└── dataset_multi_abnormality_labels_valid_predicted_labels.csv

Update the path variables at the top of src/train_grounding.py, src/train_grounding_weak.py, src/dataset_train.py, and src/dataset_valid.py if your paths differ.

Training

Fully-supervised (patch-level annotations), with selectable backbone:

python src/train_grounding.py --backbone r3d18     # R3D-18, default
python src/train_grounding.py --backbone swin3d_t  # Swin3D-T
python src/train_grounding.py --backbone mvit       # MViT-v2-S

Weakly supervised (study-level image labels only), with selectable backbone and method:

python src/train_grounding_weak.py --backbone r3d18 --method noisyor     # NoisyOR + R3D-18, default
python src/train_grounding_weak.py --backbone swin3d_t --method noisyor  # NoisyOR + Swin3D-T
python src/train_grounding_weak.py --backbone mvit --method noisyor      # NoisyOR + MViT-v2-S
python src/train_grounding_weak.py --backbone r3d18 --method gradcam     # Grad-CAM (r3d18 only)

Both scripts log to Weights & Biases by default.

---

License

The PatchChestCT annotations are released under CC BY-NC-SA 4.0, consistent with the CT-RATE Data Usage Agreement. The source CT volumes remain subject to the CT-RATE DUA and must be obtained via the official CT-RATE gated access procedure.

---