Development of a Machine Learning-Based Screening Method for Thyroid Nodules Classification by Solving the Imbalance...

Area of Science:

• Medical imaging analysis
• Machine learning in healthcare
• Oncology diagnostics

Background:

• Thyroid nodule screening often faces imbalanced datasets, where benign cases significantly outnumber malignant ones.
• This imbalance can lead to biased machine learning models, resulting in misleading assessments.
• Typical evaluation methods may not adequately capture model performance on imbalanced data.

Purpose of the Study:

• To investigate the impact of imbalanced data on machine learning models for preoperative thyroid nodule screening.
• To compare the performance of different machine learning models using both balanced and imbalanced datasets.
• To highlight the importance of appropriate evaluation metrics for reliable diagnostic tools.

Main Methods:

• A retrospective study analyzed ultrasonography features from 431 thyroid nodules across 313 patients.
• A hybrid resampling technique (Smote) was employed to create a balanced dataset.
• Support Vector Classification (SVC) and Logistic Regression models were trained and evaluated.

Main Results:

• The prevalence of malignant nodules was 14%. Microcalcification and taller-than-wide shape were key indicators of malignancy.
• Model 3 (SVC on balanced data) achieved the highest geometric mean (84.2%), outperforming models trained on unbalanced data.
• Accuracy and Area Under the Curve (AUC) remained high across models, underscoring the need for nuanced evaluation.

Conclusions:

• Data imbalance poses a significant challenge in developing accurate machine learning models for thyroid nodule screening.
• Employing data balancing techniques and appropriate evaluation metrics is essential for reliable preoperative assessments.
• This study emphasizes the need for careful consideration of data characteristics and performance measures in clinical AI applications.