Abstract
This paper presents a large-scale, publicly available dataset of 1,014 whole-body Fluorodeoxyglucose (FDG)-PET/CT studies designed to advance deep learning-based automated cancer detection. The dataset includes 501 studies of patients with malignant lymphoma, melanoma, and non-small cell lung cancer (NSCLC), alongside 513 negative control studies. All malignant lesions were manually segmented in 3D by expert radiologists. We provide anonymized DICOM files, segmentation masks, and processing scripts, demonstrating the dataset's utility through a baseline deep learning model that achieves high correlation with human expert labels.
Introduction
While deep learning has revolutionized medical imaging, its progress in PET/CT (Positron Emission Tomography/Computed Tomography) has been hampered by a significant lack of high-quality, annotated data. PET/CT is a hybrid modality where CT provides anatomical detail and PET reveals metabolic activity. A major challenge for automated systems is distinguishing malignant tumors from healthy organs—such as the brain, heart, and bladder—that naturally exhibit high metabolic activity. Without large-scale, manually annotated datasets, developing reliable algorithms for automated tumor burden measurement remains difficult.
Methodology
To address this "data desert," we curated a standardized resource of 1,014 studies acquired on a single Siemens Biograph mCT scanner between 2014 and 2018. This consistency eliminates inter-scanner variability, providing a clean benchmark for algorithm development.
Expert Annotation
An experienced radiologist with over 10 years of expertise manually segmented every FDG-avid malignant lesion identified in the clinical reports. These segmentations were performed in a slice-by-slice (3D) manner on the PET images to ensure high-fidelity ground truth labels.
Baseline Model Training
To validate the dataset, we trained a baseline model using the nnU-Net framework, a state-of-the-art 3D U-Net architecture for medical image segmentation. We employed 5-fold cross-validation to assess the model's ability to replicate expert human annotations.
Dataset Characteristics
The FDG-PET-CT-Lesions dataset is hosted on The Cancer Imaging Archive (TCIA) and includes:
- Positive Samples (501): Patients with histologically confirmed malignant lymphoma, melanoma, or NSCLC.
- Negative Controls (513): Patients without metabolically active malignant disease, essential for training models to avoid false positives in healthy tissues.
- Accessibility: Data is provided in original DICOM format with scripts for conversion to NIfTI, mha, or hdf5 formats.
Key Results
The baseline AI model demonstrated the high quality of the manual annotations through several key metrics:
- Volume Correlation: A high correlation (r = 0.85) was found between the AI’s predicted tumor volume and the manual expert volume.
- Segmentation Accuracy: The model achieved a mean Dice score of 0.73 on positive datasets.
- Clinical Robustness: The mean false positive volume was limited to 8.1 ml, indicating the model effectively ignored physiological "hot spots."
Discussion and Future Work
The primary limitation of this dataset is its single-center, single-scanner origin. While this ensures consistency, further validation on multi-vendor data is necessary for broader clinical application. Additionally, while the current focus is on three specific cancer types, the dataset provides a foundation for future expansion into other malignancies and longitudinal studies to track treatment response. This resource fills a critical gap, providing the high-quality labels needed to train the next generation of automated cancer monitoring tools.
Summary updated