Ultrasound Thyroid Nodule Segmentation Algorithm Based on DeepLabV3+ with EfficientNet.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Thyroid nodule segmentation in ultrasound images is crucial for diagnosis but challenging due to image noise and artifacts.
• Existing deep learning methods for thyroid nodule segmentation often yield suboptimal performance.

Purpose of the Study:

• To introduce and evaluate the novel application of EfficientNet-B7 as a backbone for the DeepLabV3+ architecture for thyroid nodule segmentation.
• To assess the performance of this new deep learning model on clinical and public datasets.

Main Methods:

• Implementation of the DeepLabV3+ architecture utilizing EfficientNet-B7 as its feature extraction backbone.
• Training and validation of the model on a dataset from the First Affiliated Hospital of Zhengzhou University and two public ultrasound datasets.

Main Results:

• The proposed EfficientNet-B7-DeepLabV3+ model achieved high segmentation accuracy.
• Key performance metrics include pixel accuracy (PA) of 97.67%, Dice similarity coefficient of 0.8839, and Intersection over Union (IoU) of 79.69%.

Conclusions:

• The integration of EfficientNet-B7 with DeepLabV3+ represents a significant advancement in thyroid nodule segmentation.
• This deep learning approach demonstrates superior performance compared to most traditional segmentation networks, offering a promising tool for clinical applications.