Mitigating Class Imbalance Through a Dynamic Training Regime: Application to Thyroid Nodule Classification

Abstract

Accurate classification of thyroid nodules from ultrasound images is complicated by the unequal distribution of benign and malignant cases, which can reduce the ability of deep learning models to identify clinically important abnormalities. Although a range of imbalance-handling techniques has been proposed, many depend on synthetic data generation, extensive model modifications, or computationally demanding architectures that may limit practical implementation. This research explores a training-oriented approach based on the Dynamic Balanced Training Regimes (DBTR) framework. Rather than altering the original dataset, the method combines learning from the natural class distribution with iterative training on balanced data subsets to strengthen recognition of underrepresented cases while retaining the characteristics of real clinical images. The proposed framework will be assessed using lightweight convolutional neural networks on the publicly available DDTI thyroid ultrasound dataset. The study aims to determine whether this strategy can improve classification performance under imbalanced conditions while remaining suitable for deployment in resource-constrained healthcare environments.