A Novel Multi-task Framework with Super-Resolution Directed Network for Thyroid Nodule Segmentation in Ultrasound Images

One of the first indicators of thyroid cancer is the development of an abnormal lump, or nodule, in the thyroid gland. To detect and treat thyroid cancer at an early stage, nodule detection is highly desirable. Due to speckle noise, low resolution, intensity heterogeneity, and low contrast, segmenting and detecting thyroid nodules is challenging. We present a multi-task learning framework for the segmentation of thyroid glands and nodules at the same time, using the underlying backbone networks to determine the nodule's location within the thyroid area. To begin, we build a super-resolution directed network to increase the pixel density of the original ultrasound scan. The original thyroid ultrasound image's high-frequency information can be enhanced and made more comprehensive using a super-resolution directed network. The segmentation task is then carried out using our proposed “parallel atrous convolution” (PAC) module. The proposed framework training, validation, and testing are done on the UTNI-2021 dataset. Our proposed framework outperformed most existing frameworks in terms of Precision, Recall, F-score, mIoU, and Dice with experimental data showing that it achieves 93.8%, 88.9%, 89.2%, 84.6%, and 85.4%, respectively.