Assessment of encoder-decoder-based segmentation models for thyroid ultrasound images

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Encoder-decoder models are crucial for semantic segmentation, classifying image pixels into regions of interest (ROI) or background.
• Accurate segmentation of thyroid tumor ultrasound (TTUS) images is vital for diagnosis and treatment planning.

Purpose of the Study:

• To implement and evaluate various encoder-decoder semantic segmentation models for TTUS image segmentation.
• To compare the performance of different models, including SegNet, U-Net, mobileNetv2, ResNet, Xception, and Inception networks.

Main Methods:

• Implemented several encoder-decoder architectures (SegNet, U-Net, mobileNetv2, ResNet, Xception, Inception) for TTUS image segmentation.
• Utilized transfer learning to train models on original and despeckled TTUS images.
• Evaluated model performance using mean Intersection over Union (mIoU) and mean Dice Coefficient (mDC) metrics.

Main Results:

• The ResNet50-based model demonstrated superior performance, achieving an mIoU of 0.87 and an mDC of 0.94.
• Both objective metrics and subjective assessment by radiologists confirmed the effectiveness of the ResNet50 model.
• The model accurately segmented lesion characteristics such as shape, margin, and echogenicity.

Conclusions:

• The ResNet50 semantic segmentation model offers highly accurate and reliable performance for TTUS image analysis.
• This model shows significant potential for integration into healthcare systems for real-time thyroid nodule identification.
• The findings support the use of advanced deep learning techniques for improved medical image analysis.