Hybrid deep feature and machine learning framework for classification of thyroid nodules in ultrasound images

Accurate differentiation between benign and malignant thyroid nodules is essential for reducing unnecessary biopsies and improving early clinical decision-making. This study aims to enhance the reliability of ultrasound-based thyroid nodule assessment by proposing an optimized computer-aided diagnosis (CAD) framework.

A hybrid CAD framework combining deep transfer learning with gradient-boosted decision tree classification was developed. High-level semantic features were extracted from a pretrained ResNet50 model and encoded as fixed-length representations. These features were then fed into the CatBoost algorithm for downstream classification, enabling the framework to leverage both the representational richness of deep neural networks and the decision efficiency of boosted tree ensembles, especially under the condition of limited annotated medical data. A comprehensive set of complementary evaluation measures was used to assess the diagnostic performance, which reflects overall accuracy, error distribution, and the model's ability to distinguish clinically meaningful malignant cases.

Experimental results demonstrate that the proposed framework, which couples transfer-learned deep features with CatBoost, achieves superior discrimination between benign and malignant thyroid nodules compared with conventional approaches. It exhibits improved robustness across variations in ultrasound appearance and maintains stable performance without the need for extensive parameter tuning.

These findings highlight the potential of the proposed method as an efficient and reliable tool for computer-aided thyroid nodule diagnosis. They also underscore the framework's suitability for integration into real-world clinical workflows, which could further optimize clinical decision-making and reduce unnecessary medical interventions.