GDSSA-Net: A gradually deeply supervised self-ensemble attention network for IoMT-integrated thyroid nodule segmentation

Abstract

The integration of deep learning techniques in the Internet of Medical Things (IoMT) has significantly advanced the early detection of life-threatening diseases such as thyroid cancer, one of the most lethal tumors. Accurate delineation of thyroid nodules in ultrasound images is essential for timely diagnosis and for effective treatment. This research introduces a novel deep-learning framework tailored for IoMT environments, aimed at the automatic segmentation of thyroid nodules in ultrasound images. We propose a Gradually Deeply Supervised Self-ensemble Attention Network (GDSSA-Net), which employs encoder to extract features from sonographic scans and integrates a gated attention mechanism within the decoder to refine features while filtering out irrelevant information. To enhance the learning process, we developed a novel Gradual Deep Supervision (GDS) strategy, utilizing three variations of ground truth to deeply supervise the network. Additionally, our approach employs self-ensembling mechanisms by ensembling outputs of the shallower branches alongside the main branch to improve the thyroid nodule segmentation. To validate the superiority and generalizability of GDSSA-Net, we conducted extensive evaluations on two publicly available datasets, DDTI and TN3K. Experimental results demonstrate that our method surpasses its simplified variants and existing state-of-the-art models in terms of quantitative metrics and qualitative assessments. Specifically, our model achieves a Dice coefficient of 79.85% and 84.27% on DDTI and TN3K, respectively. The source code for our proposed model is publicly available at https://github.com/harisghafoor/GDSSA-Net.