Learning vector quantizer in the investigation of thyroid lesions

Area of Science:

• Computational pathology
• Artificial intelligence in medicine
• Thyroid cancer diagnostics

Background:

• Fine needle aspiration (FNA) is crucial for thyroid nodule evaluation.
• Accurate differentiation of benign from malignant thyroid lesions is essential for patient management.
• Image analysis and machine learning offer potential for improving diagnostic accuracy.

Purpose of the Study:

• To evaluate the effectiveness of learning vector quantization (LVQ) neural networks for classifying thyroid lesions.
• To assess the capability of LVQ in discriminating between benign and malignant thyroid tumors using FNA cytology.
• To explore the utility of image morphometry features in conjunction with LVQ for thyroid lesion diagnosis.

Main Methods:

• Cytologic smears from 198 thyroid lesions were analyzed.
• 25 nuclear features (size, shape, texture) were extracted using a custom image analysis system.
• Two LVQ classifiers were trained on approximately 30% of cases and tested on the remaining 139 cases.

Main Results:

• LVQ achieved high accuracy (97.8% overall accuracy) in discriminating between benign and malignant thyroid lesions.
• The LVQ classifiers demonstrated robust performance in the binary classification task.
• Reliable discrimination between specific cytologic types of thyroid lesions was not achieved.

Conclusions:

• LVQ neural networks combined with image morphometry show promise for improving diagnostic accuracy in thyroid FNA.
• This approach may be particularly beneficial for ambiguous cases, such as suspicious follicular neoplasms and oncocytic tumors.
• The findings suggest a potential role for AI-driven image analysis in enhancing thyroid cancer diagnosis.