Self-supervised enhanced thyroid nodule detection in ultrasound examination video sequences with multi-perspective...

Area of Science:

• Medical imaging
• Artificial intelligence
• Ultrasound technology

Background:

• Thyroid nodule detection via ultrasound is crucial but manual methods are time-consuming and subjective.
• Current machine learning approaches for ultrasound images are limited by low signal-to-noise ratio and poor tissue contrast.

Purpose of the Study:

• To develop a video-based deep learning model for accurate and real-time thyroid nodule detection.
• To leverage temporal context in ultrasound videos to overcome limitations of image-based methods.
• To introduce a self-supervised learning method to reduce the need for extensive medical image annotation.

Main Methods:

• A video-based deep learning model with adjacent frame perception (AFP) was proposed to aggregate contextual features from ultrasound videos.
• A patch scale self-supervised model (PASS) was developed and trained on unlabeled data to enhance the AFP model's performance.
• The models were trained and evaluated using a dataset of 92 ultrasound videos (23,773 frames).

Main Results:

• The AFP model improved the average precision at 50% intersection-over-union (AP@50) from 0.256 to 0.390.
• The PASS-enhanced AFP model further boosted AP@50 to 0.425.
• Both AFP and PASS demonstrated performance improvements across various evaluation perspectives.

Conclusions:

• The proposed video-based model effectively mitigates challenges associated with low signal-to-noise ratio and tissue contrast in ultrasound imaging.
• The integration of adjacent frame perception and self-supervised learning enables accurate, real-time thyroid nodule detection.
• Ablation experiments confirm the efficacy of both the AFP and PASS models for enhanced thyroid nodule identification.