Integrating User-Input into Deep Convolutional Neural Networks for Thyroid Nodule Segmentation.

Area of Science:

• Medical Imaging
• Computer-Aided Diagnosis
• Ultrasound Technology

Background:

• Accurate thyroid nodule boundary delineation is crucial for cancer risk assessment and nodule categorization.
• Current manual or bounding-box methods for thyroid nodule assessment are subjective, leading to poor inter-observer agreement.
• Computer-aided diagnosis (CAD) systems can enhance consistency and precision in nodule assessment.

Purpose of the Study:

• To present a novel, user-friendly approach for effective thyroid nodule segmentation and tracking in ultrasound images.
• To develop a system that predicts nodule segmentation across an entire ultrasound sweep from a single user click.
• To reduce subjectivity and improve efficiency in thyroid nodule assessment.

Main Methods:

• A novel computer-aided diagnosis model was developed for thyroid nodule segmentation.
• The system utilizes a single user click on the region of interest within an ultrasound sweep.
• The model predicts nodule segmentation for the entire sequence of frames.

Main Results:

• The proposed one-click segmentation method significantly outperforms the bounding box approach.
• Quantitative evaluations demonstrated superior performance using the dice score on a dataset of 372 ultrasound images.
• The approach reduces expert time and potential variability in thyroid nodule assessment.

Conclusions:

• The developed one-click approach offers an effective and efficient solution for thyroid nodule segmentation in ultrasound imaging.
• This method minimizes user interaction, enabling precise boundary identification for volumetric measurement and characterization.
• The approach has potential for rapid labeling of large datasets for training machine learning algorithms in thyroid imaging.