Large-scale generative tumor synthesis in computed tomography images for improving tumor recognition

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Radiology

Background:

• AI-driven tumor recognition shows promise for precise screening and diagnosis.
• Progress is limited by the scarcity of annotated medical imaging datasets.
• Manual annotation by radiologists is time-consuming and resource-intensive.

Purpose of the Study:

• To introduce FreeTumor, a Generative AI framework for large-scale tumor synthesis.
• To mitigate data scarcity in medical imaging datasets for AI training.
• To augment existing datasets with realistic synthetic tumors for improved AI model performance.

Main Methods:

• Developed FreeTumor, a Generative AI framework leveraging limited labeled and large-scale unlabeled data.
• Synthesized a large number of realistic tumors to augment training datasets.
• Curated a dataset of 161,310 Computed Tomography (CT) volumes, with only 2.3% annotated.

Main Results:

• FreeTumor successfully synthesized a large volume of realistic tumors.
• Synthetic tumor quality was rigorously validated by 13 board-certified radiologists.
• AI models trained with FreeTumor-generated data demonstrated superior tumor recognition compared to state-of-the-art methods.

Conclusions:

• FreeTumor effectively addresses data scarcity in medical imaging through high-quality tumor synthesis.
• The framework shows significant potential for improving AI-driven tumor screening and diagnosis.
• Generated synthetic tumors enhance the performance of tumor recognition models, indicating promising clinical applications.