Bone metastasis scintigram generation using generative adversarial learning with multi-receptive field learning and...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Biology

Background:

• Deep learning models for SPECT bone scintigram analysis require large datasets.
• Limited data availability hinders the development of high-performing deep learning models.
• Automated data generation is crucial for expanding SPECT bone scintigram datasets.

Purpose of the Study:

• Introduce a deep learning-based generation model (BMS-Gen) for creating realistic SPECT bone scintigram samples.
• Generate novel, non-identical samples to augment existing datasets.
• Facilitate the development of advanced automated analysis tools.

Main Methods:

• Utilize a generative adversarial learning architecture for BMS-Gen.
• Employ multi-receptive field learning with multiple input conditions for realism.
• Incorporate generative adversarial learning to maintain sample diversity.
• Implement a two-stage training strategy to enhance generated sample quality.

Main Results:

• BMS-Gen achieved superior performance with FID, MSE, and PSNR scores of 1678.0, 69.33, and 19.51, respectively.
• Augmenting datasets with BMS-Gen samples improved F-1 scores by up to 3.01% for classification.
• DSC scores for segmentation tasks increased by up to 6.83% with generated samples.

Conclusions:

• BMS-Gen serves as a valuable tool for augmenting bone scintigram data.
• The model significantly aids the development of deep learning-based automated analysis of SPECT bone scintigrams.
• This approach promises to enhance the accuracy and efficiency of diagnostic imaging.