Radiomics encompasses a variety of image attributes, including intensity, texture, shape, and spatial relationships among pixels. In this study, we applied four feature selection techniques: Principal Component Analysis (PCA), Linear Discriminant Analysis (LDA), t-Distributed Stochastic Neighbor Embedding (t-SNE), and Recursive Feature Elimination (RFE), to reduce the dimensionality of radiomic features extracted from thyroid thermal images. Subsequently, we employed three models: U-Net with a VGG16 backbone, U-Net with ResNet50 backbone, and U-Net with DenseNet121 backbone to segment thyroid nodules. The U-Net with a VGG16 backbone combined with the LDA model achieved an average Dice coefficient of 0.4035, significantly outperforming the other models. These results highlight the potential of feature reduction techniques in enhancing thyroid nodule segmentation in thermal imaging.