Automatic Classification of Nodules from 2D Ultrasound Images Using Deep Learning Networks.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Thyroid nodule evaluation relies on physician interpretation of 2D ultrasound images.
• Visual assessment can lead to unnecessary fine needle aspirations (FNA).
• Accurate classification is crucial for patient management.

Purpose of the Study:

• To develop an automated thyroid ultrasound image classification system.
• To prevent unnecessary fine needle aspirations (FNA) in clinical practice.
• To improve diagnostic accuracy for thyroid nodules.

Main Methods:

• A deep learning model (DenseNet with attention) was fine-tuned for classification.
• A dataset of 591 thyroid nodule ultrasound images was used.
• Techniques included data augmentation, class weighting, and Grad-CAM for interpretability.

Main Results:

• The system achieved high performance: 0.94 accuracy, 0.93 F1-score, and 0.96 sensitivity.
• Gradient-weighted class activation maps (Grad-CAM) provided decision-making insights.
• The model demonstrated reliability for clinical end-users.

Conclusions:

• A deep learning architecture effectively classifies thyroid nodules for FNA necessity.
• The system shows potential to reduce unnecessary FNAs.
• High accuracy and minimal false negatives support clinical utility.