Morphological and wavelet features towards sonographic thyroid nodules evaluation.

Area of Science:

• Medical Imaging
• Computer-Aided Diagnosis
• Biomedical Signal Processing

Background:

• Thyroid nodules are common, and accurately assessing malignancy risk is crucial for patient management.
• Ultrasonography is a primary imaging modality for thyroid nodule evaluation.
• Distinguishing benign from malignant thyroid nodules can be challenging, necessitating advanced diagnostic tools.

Purpose of the Study:

• To develop and evaluate a computer-based classification scheme for thyroid nodule malignancy risk using morphological and wavelet-based features.
• To assess the performance of support vector machines (SVM) and probabilistic neural networks (PNN) in classifying thyroid nodules.
• To investigate the impact of speckle noise on classification accuracy.

Main Methods:

• A dataset of 85 cytologically confirmed thyroid ultrasound images (54 low-risk, 31 high-risk) was used.
• Twenty features were extracted: 12 shape-based and 8 wavelet-based (local maxima).
• Support vector machines and probabilistic neural networks were employed as classifiers, with and without speckle noise considerations.

Main Results:

• Without speckle, SVM achieved an area under the ROC curve (AUC) of 0.96, and PNN achieved 0.91.
• With speckle, SVM's AUC decreased to 0.88, and PNN's to 0.86.
• The proposed features demonstrated strong differentiation power for both classifiers.

Conclusions:

• The developed computer-based classification scheme, utilizing novel wavelet features, shows high potential for accurate thyroid nodule malignancy risk evaluation.
• The proposed features can enhance diagnostic accuracy and reduce misdiagnosis rates in thyroid cancer screening.
• While speckle noise impacts performance, the system remains effective, suggesting robustness for clinical application.