Human-machine collaborative risk assessment model for thyroid nodules based on local attention and multi-scale...

Area of Science:

• Artificial Intelligence in Medicine
• Medical Imaging Analysis
• Diagnostic Decision Support

Background:

• Thyroid nodules are common, requiring accurate risk assessment for appropriate management.
• Radiologist experience significantly impacts diagnostic accuracy for thyroid nodules.
• Current diagnostic methods can be subjective and vary in performance.

Purpose of the Study:

• To evaluate a novel human-machine collaborative risk assessment model for thyroid nodules.
• To compare the model's performance against radiologists with diverse experience levels.
• To assess the impact of an assistive strategy using the model on diagnostic outcomes.

Main Methods:

• A multi-center study utilized ultrasound images from 8063 patients for model training and validation.
• A deep learning model with local attention and multi-scale features was developed.
• The model's diagnostic performance was benchmarked against junior, intermediate, and senior radiologists.
• An assistive strategy allowed radiologists to refine diagnoses based on model outputs.

Main Results:

• The model achieved high accuracies in identifying specific thyroid nodule features (e.g., composition 0.966, echogenicity 0.809).
• The model's area under the receiver operating characteristic curve (AUROC) for nodule diagnosis was 0.882, outperforming junior radiologists (0.789).
• The assistive strategy significantly improved the junior radiologist's AUROC to 0.859 and increased sensitivity from 66.11% to 80.00%.

Conclusions:

• The human-machine collaborative model accurately identifies thyroid nodule risk features.
• The model effectively enhances diagnostic performance, particularly for less experienced radiologists.
• The assistive strategy shows promise in improving diagnostic accuracy and sensitivity for thyroid nodule detection.