On the identification of thyroid nodules using semi-supervised deep learning.

Area of Science:

• Medical imaging analysis
• Machine learning in healthcare
• Oncology

Background:

• Early detection of malignant thyroid nodules is critical but challenging due to their rarity.
• Traditional rule-based and expert-based methods struggle with identifying these rare cases.
• Machine learning (ML) offers a promising approach to improve diagnostic accuracy for rare diseases.

Purpose of the Study:

• To investigate the application of ML for detecting rare malignant thyroid nodules.
• To evaluate the effectiveness of an auto-encoder based neural network model for this task.
• To compare ML performance against traditional classification algorithms.

Main Methods:

• Utilized a dataset of ultrasound (US) texture features and assessment scores from 636 patients.
• Incorporated data from various health institutions' assessment methods (e.g., Korean, American, European thyroid societies).
• Employed an auto-encoder based neural network for detecting rare malignant cases.

Main Results:

• The auto-encoder model achieved an average Recall score of 0.98 for malignant cases.
• The model demonstrated a Sensitivity score of 1.00 for both malignant and non-malignant cases.
• Performance surpassed traditional classification algorithms, even after Synthetic Minority Oversampling Technique (SMOTE) oversampling.

Conclusions:

• Auto-encoder based neural networks are highly effective for detecting rare malignant thyroid nodules.
• ML models significantly improve upon traditional methods in identifying rare thyroid cancer cases.
• This approach holds potential for enhancing early cancer detection in healthcare settings.