MELA: A Benchmark for Mediastinal Lesion Analysis in CT Scans

A benchmark dataset of 1,100 CT scans collected from patients with lesions in the mediastinum. It was created by Shuang Song of Wuhan University to facilitate research into automatic lesion detection and diagnosis. The dataset is split into training, validation, and test subsets, with this part containing 250 training CTs.

Use Cases

• Training deep learning models for mediastinal lesion detection based on annotated CT scans.
• Benchmarking algorithm performance for medical image segmentation tasks.
• Developing diagnostic support tools for radiologists analyzing lesions in the chest.

Strengths

• Contains 1,100 CT scans, providing a substantial volume of medical imaging data.
• Explicitly structured for ML benchmarking with defined training (770 scans), validation (110 scans), and test (220 scans) splits.
• Data is provided in the NIfTI format (nii.gz), a standard for medical imaging analysis.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Last update date is unknown; freshness unverified.
• Geographic and demographic provenance of the patient scans is not specified, which may limit assessment of potential bias.

Provenance

Source

Wuhan University

Collection Method

Collected from patients with one or more lesions in the mediastinum.