A knowledge-interpretable multi-task learning framework for automated thyroid nodule diagnosis in ultrasound videos

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Ultrasound is the primary method for thyroid nodule diagnosis due to its accessibility and safety.
• Computer-aided diagnosis (CAD) for thyroid nodules is gaining traction but faces limitations with static images or opaque models.
• Existing CAD systems often struggle with interpretability and generalizability in real-world diagnostic workflows.

Purpose of the Study:

• To develop a user-friendly, interpretable CAD framework for automated thyroid nodule diagnosis from ultrasound videos.
• To simulate a radiologist's diagnostic workflow for enhanced accuracy and clinical relevance.
• To improve the efficiency and generalizability of thyroid nodule classification using temporal video analysis.

Main Methods:

• A two-part framework was developed: image-based interpretation for TI-RADS scores and temporal modeling of video sequences.
• Frame-wise image analysis incorporated prior knowledge to generate embedded representations.
• A sequence modeling approach selectively enhanced temporal information for improved classification of nodule malignancy.

Main Results:

• The proposed framework demonstrated superior performance compared to existing state-of-the-art video classification methods.
• The system provided interpretable results, aiding radiologists in understanding diagnostic predictions.
• The approach showed enhanced efficiency and generalizability in classifying thyroid nodules.

Conclusions:

• The developed CAD framework offers a promising tool for automated thyroid nodule diagnosis in ultrasound videos.
• This technology can assist radiologists, reduce diagnostic workload, and potentially improve patient care.
• The interpretable nature of the framework enhances trust and clinical adoption for AI in thyroid imaging.