BS-80K: The first large open-access dataset of bone scan images

Area of Science:

• Medical Imaging
• Machine Learning
• Oncology

Background:

• Radionuclide bone scanning is crucial for detecting bone metastasis.
• Manual analysis of bone scans is time-consuming and subjective.
• Existing machine learning datasets for bone scans are often small and private, hindering research.

Purpose of the Study:

• To introduce BS-80K, a large, publicly available dataset of bone scan images.
• To establish a benchmark for computer-aided diagnosis of bone metastasis using deep learning models.
• To facilitate further research in automated bone scan analysis.

Main Methods:

• A dataset of 82,544 bone scan images (BS-80K) from 3,247 patients was curated.
• Images include anterior and posterior views with region-wise slices and bounding box annotations.
• Six popular deep learning models were benchmarked for classification and object detection tasks.

Main Results:

• Deep learning models achieved high accuracy and specificity (around 95%) for metastasis prediction.
• Object detection models showed competitive performance with average precision ranging from 0.1334 to 0.2484.
• Models trained on BS-80K significantly outperformed those trained on smaller datasets.

Conclusions:

• BS-80K is the first large, public bone scan dataset, supporting diverse research in computer-aided diagnosis.
• The dataset's scale significantly enhances the performance of machine learning models for bone metastasis detection.
• BS-80K is expected to accelerate advancements in automated analysis of bone scan images.