Computer aided diagnosis of thyroid nodules based on the devised small-datasets multi-view ensemble learning.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Computational Biology

Background:

• Deep learning models require large datasets for effective training, posing a challenge for medical image analysis where data is often scarce.
• Diagnosing benign and malignant thyroid nodules accurately is crucial for patient outcomes.

Purpose of the Study:

• To develop a robust deep learning diagnostic model for thyroid nodules using small ultrasound image datasets.
• To enhance the performance of diagnostic models by integrating multi-view information and ensemble learning.

Main Methods:

• A multi-view ensemble learning strategy was proposed, combining results from GoogleNet trained on ultrasound images, U-Net extracted medical features, and mRMR selected statistical/texture features.
• An Xgboost classifier was used to analyze feature contributions and generate intermediate diagnostic results.
• A majority voting mechanism was employed to achieve the final diagnosis, with the method also applied to sequence ultrasound images.

Main Results:

• The proposed method significantly improved diagnostic accuracy compared to traditional deep learning models trained on small datasets.
• Accuracies of 92.11% and 92.54% were achieved on normal and sequence image datasets, respectively.
• The ensemble approach effectively addressed challenges posed by limited data and slight variations in sequence images.

Conclusions:

• Multi-view ensemble learning offers a powerful solution for training effective deep learning diagnostic models with limited medical image data.
• The proposed method demonstrates superior performance in thyroid nodule diagnosis, highlighting its clinical potential.