A Comparative Study on Thyroid Segmentation Using Filters and Transfer Learning

Abstract

Ultrasound (US) images are commonly used for analyzing soft tissues because they are non-invasive and low-cost. US image quality can be compromised by low contrast and noise, potentially impairing radiologists' interpretation and affecting the accuracy of computer algorithms in segmentation and analysis characterization. Successful segmentation models are often computationally intensive, making them less practical for diagnostic applications. In this work, a U-Net model was developed that employs a CNN as the encoder for segmentation, combined with image preprocessing using border-detection filters to enhance edges and reduce noise. We compare the performance of four segmentation models, all implemented with the same data. Notably, U-Net with Xception model encoder and preprocessing with bilateral filter and CLAHE achieves high performance, with a Dice score of 0.7476 and an IoU of 0.6862.