Improving AI models for rare thyroid cancer subtype by text guided diffusion models.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Artificial intelligence (AI) in oncology imaging faces challenges in diagnosing rare tumors.
• Scarce data for uncommon thyroid cancer subtypes leads to misdiagnosis in ultrasound imaging.
• Existing data augmentation methods fail to capture unique disease variations, limiting AI model performance.

Purpose of the Study:

• To develop a novel text-driven generative method for improving the detection of rare tumors in medical imaging.
• To enhance the accuracy and robustness of AI models for diagnosing uncommon cancer subtypes.

Main Methods:

• A text-driven generative approach was proposed, integrating clinical insights with image generation.
• Synthetic ultrasound images reflecting rare thyroid cancer subtypes were created.
• The method was evaluated on its ability to improve diagnostic metrics and generalizability.

Main Results:

• The proposed method demonstrated substantial gains in diagnostic metrics for rare tumor detection.
• Generated synthetic samples showed high authenticity and diversity compared to existing methods.
• The approach generalized effectively to diverse private and public datasets with various rare cancers.

Conclusions:

• Text-guided image augmentation significantly enhances AI model accuracy and robustness for rare tumor detection.
• This approach offers a promising solution for improving diagnostic reliability in clinical oncology.
• The method facilitates more reliable and widespread clinical adoption of AI in cancer imaging.