Thyroid Nodules Classification using Weighted Average Ensemble and D-CRITIC based TOPSIS Methods for Ultrasound...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Thyroid disorders are common, necessitating early detection for better patient outcomes.
• Traditional methods for thyroid nodule analysis are time-consuming and complex.
• Ultrasound imaging is crucial for timely detection of malignant thyroid nodules.

Purpose of the Study:

• To develop computer-aided diagnosis tools for early malignant thyroid nodule detection using ultrasound images.
• To address challenges of small medical datasets and class imbalance in machine learning models.
• To design and benchmark ensemble learning models for superior diagnostic accuracy.

Main Methods:

• Investigated four recent image transformer and mixer models for thyroid nodule detection.
• Developed weighted average ensemble models with weights optimized by the Hunger Games Search (HGS) algorithm.
• Utilized the D-CRITIC-TOPSIS method for ranking the performance of different ensemble models.

Main Results:

• The gMLP+ViT ensemble model achieved 89.70% accuracy with an 80:20 data split.
• The gMLP+FNet+Mixer-MLP ensemble model achieved 82.18% accuracy with a 70:30 data split.
• The proposed ensemble models demonstrated superior performance on an imbalanced thyroid ultrasound dataset.

Conclusions:

• Ensemble learning models significantly improve thyroid nodule detection accuracy compared to individual models.
• The developed models offer a promising approach for computer-aided diagnosis of thyroid cancer.
• This research highlights the effectiveness of ensemble methods in handling imbalanced medical imaging datasets.