Adaptive Control of Data Augmentation for Thyroid Ultrasound Image Segmentation

Abstract

The aim of this study was to investigate the influence of the proportion of real and augmented data on the performance of thyroid nodule segmentation in ultrasound images. For this purpose, a parameter R was introduced, defined as the ratio of the number of augmented images to the number of original images, and its effect on segmentation quality was analyzed for different encoder architectures. The study considered the formation of training datasets with varying shares of augmented data, followed by model training and evaluation of segmentation results on the DDTI dataset. Segmentation performance was assessed using standard evaluation metrics, namely the Dice coefficient and Intersection over Union (IoU). The results show that the dependence of segmentation quality on the parameter R is non-monotonic: there exists an optimal value 𝑅_maxat which the best performance is achieved, as well as a region 𝑅_minwhere the segmentation quality deteriorates. It is demonstrated that the ViT-Base architecture is the most sensitive to excessive augmentation, whereas ConvNeXt-Tiny and Swin Transformer-Tiny exhibit greater robustness to changes in the proportion of synthetic data.