Thyroid Nodule Detection and Classification on Small Datasets: An Ensemble Deep Learning Approach with Attention...

Area of Science:

• Medical Imaging and Diagnostics
• Artificial Intelligence in Healthcare
• Deep Learning for Medical Analysis

Background:

• Thyroid nodule classification on ultrasound is challenging due to limited labeled data and significant class imbalance.
• Existing methods struggle with the complexities of accurately differentiating benign from malignant thyroid nodules.

Purpose of the Study:

• To develop and evaluate an integrated deep learning framework for improved thyroid nodule classification.
• To address the challenges of limited data and class imbalance in thyroid ultrasound image analysis.

Main Methods:

• An integrated framework combining YOLO-based region-of-interest detection and an enhanced ResNet18 classifier was proposed.
• A dataset of 522 thyroid ultrasound images was utilized, with strategic data augmentation, focal loss, and ensemble learning.
• ResNet18 was enhanced with a convolutional block attention module, and training incorporated techniques like mixup and cosine annealing.

Main Results:

• The ensemble model achieved 85.4% accuracy, 86.4% sensitivity, and 84.2% specificity on an independent test set.
• Validation across internal and external datasets demonstrated robust diagnostic performance, with accuracy ranging from 77.8% to 85.7%.
• The findings indicate that advanced regularization and ensemble learning enhance generalizability, even with small medical datasets.

Conclusions:

• A lightweight ResNet18 architecture with strategic optimization outperforms deeper networks on small medical datasets for thyroid nodule classification.
• The proposed deep learning framework shows promising diagnostic performance across multiple validation cohorts.
• This framework represents a valuable computer-aided diagnosis tool for thyroid nodule assessment.