Doppler-Enhanced Deep Learning: Improving Thyroid Nodule Segmentation with YOLOv5 Instance Segmentation

Mahmoud El Hussieni · 2025

Read Doppler-Enhanced Deep Learning: Improving Thyroid Nodule Segmentation with YOLOv5 Instance Segmentation by Mahmoud El Hussieni (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2512-00639.txt.

Introduction

Thyroid nodules are common, and telling benign from malignant ones early guides treatment. Mahmoud El Hussieni presents a Doppler-enhanced deep learning approach to thyroid nodule segmentation in 2025, using YOLOv5 instance segmentation on ultrasound images. Rising thyroid cancer prevalence has spurred computer-aided detection methods.

Accurate segmentation of nodules is a critical first step toward AI-assisted clinical decision support. Doppler imaging adds blood-flow information that ordinary ultrasound lacks, which may sharpen the nodule boundaries the model must find. Better segmentation supports radiologists measuring and tracking nodules across visits. Instance segmentation draws each nodule separately, not just a category mask. Ultrasound is cheap and radiation-free, which is why better analysis of it matters. For scan teams, the gain is time. A tool that draws each lump with care can cut the hours spent on marks, and leave more hours for hard calls.

Catalog record

CREATORMahmoud El Hussieni
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE25 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2512-00639

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Doppler-Enhanced Deep Learning: Improving Thyroid Nodule Segmentation with YOLOv5 Instance Segmentation Mahmoud El Hussieni Abstract The increasing prevalence of thyroid cancer globally has led to the development of various computer-aided detection methods. Accurate segmentation of thyroid nodules is a critical first step in the development of AI-assisted clinical decision support systems. This study focuses on instance segmentation of thyroid nodules using YOLOv5 algorithms on ultrasound images. We evaluated multiple YOLOv5 variants (Nano, Small, Medium, Large, and XLarge) across two dataset versions, with and without doppler images. The YOLOv5-Large algorithm achieved the highest performance with a dice score of 91% and mAP of 0.87 on the dataset including doppler images. Notably, our results demonstrate that doppler images, typically excluded by physicians, can significantly improve segmentation performance. The YOLOv5-Small model achieved 79% dice score when doppler images were excluded, while including them improved performance across all model variants. These findings suggest that instance segmentation with YOLOv5 provides an effective real-time approach for thyroid nodule detection, with potential clinical applications in automated diagnostic systems. This research was conducted in 2023 as part of the TRAICK project ( https://traick.ai/ ) with supervision of Assistant Professor: Burcu Bektas Gunes 1 Introduction The thyroid is part of the endocrine system located in the front of the throat region that produces, stores, and secretes hormones. During routine checks or based on patient complaints, abnormal nodules of different structures and sizes can sometimes form within the thyroid tissue, and these nodules can lead to cancer in advanced stages. Thyroid nodules are a common medical condition affecting a significant portion of the population, with prevalence rates ranging from 19% to 68% [ 3 ] . The most useful imaging technique for thyroid examination is ultrasonography. According to epidemiological studies, thyroid nodules are frequently detected in up to 70% of ultrasound examinations; however, only 5-15% of sonographically detected nodules are malignant [ 4 ] . Therefore, the primary clinical challenge is to reliably distinguish malignant nodules that require surgical treatment from the majority of benign nodules that do not necessitate surgery. To reduce unnecessary biopsies while detecting clinically significant cancers, various risk classification systems, such as the American College of Radiology (ACR) Thyroid Imaging Reporting and Data System (TIRADS), have been proposed [ 10 ] . While TIRADS systems have reduced biopsy rates and increased specificity, a significant number of false-positive biopsies, reported at 49-56%, persist [ 6 , 7 , 11 ] . Additionally, these systems rely on the interpretation of qualitative imaging features and are subject to intra- and inter-observer variability even among expert radiologists (Kappa=0.519) [ 7 ] . The subjectivity and variability inherent in human interpretation have prompted the exploration of AI as a potential solution to enhance the accuracy and consistency of thyroid nodule risk assessment. The accurate segmentation of thyroid nodules is a crucial first step in the development of AI-powered clinical decision support systems for the detection and diagnosis of thyroid cancer. This study analyzes the performance of instance segmentation algorithms for the automatic segmentation of thyroid nodules using YOLOv5 models. We present results across multiple YOLOv5 variants on a unique thyroid ultrasound dataset, demonstrating that real-time instance segmentation can achieve high accuracy while being computationally efficient. 2 Related Work 2.1 Deep Learning for Medical Image Segmentation Recent advances in deep learning have shown considerable potential in the segmentation of medical ultrasound images [ 2 ] . Traditional approaches for thyroid nodule segmentation relied on manual feature engineering and classical image processing techniques, which often struggled with the inherent noise and variability in ultrasound images. The introduction of deep convolutional neural networks (CNNs) has revolutionized medical image analysis. U-Net and its variants have become the de facto standard for medical image segmentation tasks due to their encoder-decoder architecture with skip connections. However, these models typically focus on semantic segmentation and may not distinguish between individual instances of the same class. 2.2 Instance Segmentation Approaches Instance segmentation extends semantic segmentation by identifying and delineating individual object instances. Mask R-CNN has been widely adopted for instance segmentation tasks, combining object detection with pixel-level segmentation masks. Abdolali et al. developed a deep learning framework based on the multi-task Mask R-CNN model with a regularization loss function that prioritizes detection over segmentation. Their experimental results indicate that the proposed method outperforms Faster R-CNN and Mask R-CNN in thyroid nodule detection [ 1 ] . 2.3 YOLO-based Approaches The You Only Look Once (YOLO) family of models represents a different paradigm in object detection, offering single-stage detection with real-time performance. YOLOv5 is a state-of-the-art single-stage object detection network, known for being lightweight and efficient. Its compact size makes it easier to train and quicker to generate predictions, making it particularly suitable for real-time applications [ 8 , 12 ] . Recent work by Yang et al. demonstrated the effectiveness of improved YOLOv5 models for thyroid nodule detection, achieving competitive results with reduced computational requirements [ 12 ] . The goal of instance segmentation is to accurately identify and outline each instance of a class within an image [ 9 ] . While YOLOv5 was initially designed for object detection, its extension to instance segmentation (YOLOv5-seg) combines the speed advantages of YOLO with pixel-level segmentation capabilities. 2.4 Research Gap While several studies have explored thyroid nodule detection using various deep learning approaches, systematic comparison of different YOLOv5 variants for instance segmentation of thyroid nodules, particularly with consideration of doppler imaging, remains limited. Most existing studies focus on either detection or semantic segmentation, and few investigate the utility of doppler images, which are often excluded from clinical evaluation. This study addresses this gap by providing comprehensive evaluation of YOLOv5 models across multiple configurations and dataset compositions. 3 Materials and Methods 3.1 Dataset With the approval (number 2022/0503) of the Ethics Committee of Istanbul Medeniyet University on 07.09.2022, images of registered patients who underwent thyroid biopsy between 2018-2020 were retrieved from the picture archiving and communication systems (PACS) of Istanbul Medeniyet University Göztepe Prof. Dr. Süleyman Yalçın City Hospital. The images were labeled by a radiology specialist with 10 years of head and neck radiology experience on the ango.ai platform. During this labeling process, the TIRADS classification and nodule shape characteristics were labeled. The ango.ai platform provides a JSON output format for the labeled images, which was processed using Python scripts to create images and masks for segmentation in YOLO and COCO formats. Some images were labeled as invalid or containing artifacts by physicians. These images (n=829) were excluded from the study, creating a dataset labeled as Version-1 (V1). The suitability of V1 images for artificial intelligence was analyzed by developers. In 197 images, the doppler-elastography feature was detected. These images were identified and excluded to create Version-2 (V2). The exclusion of these 197 images when creating V2 was done intentionally to compare the model’s performance with and without doppler images. Version-1 (V1): 2,075 patients, 4,129 images, 4,407 nodules (includes doppler images) Version-2 (V2): 1,969 patients, 3,932 images, 4,192 nodules (excludes doppler images) During ultrasound imaging, a patient may have multiple images, and each image can contain more than one nodule. In this study, since the segmentation process is performed for each nodule, each nodule was labeled individually and image masks

Read the complete text