Ultrasound Image-Based Convolutional Neural Network for Predicting Thyroid Nodule FNAB Pathological TBSRTC Class I
Thyroid nodules are common tumors in the endocrine system, with most of them being benign. Fine needle aspiration biopsy (FNAB) guided by ultrasound is the preferred method to determine the nature of thyroid nodules, and the Bethesda System of thyroid cytopathology (TBSRTC) is commonly used for diagnostic reports. However, TBSRTC class I specimens are often insufficient for accurate diagnosis, leading to repeat punctures after three months. This study aims to develop a convolutional neural network (CNN) model based on ultrasound images to predict TBSRTC class I pathology in thyroid nodules before puncture, providing an objective and efficient AI tool for physicians and reducing the proportion of class I nodules.
The study consists of two parts. In the first part, the researchers focus on developing a deep learning model to accurately identify the position of the puncture needle in thyroid nodule biopsies using ultrasound images. Two datasets are used for this analysis: dataset 1 includes cases from the Department of Ultrasound at West China Hospital of Sichuan University, and dataset 2 is a publicly available thyroid nodule dataset. The segmentation network and classification network are combined to create a comprehensive model for needle position classification. The performance of multiple models, including resnet-50, Vit base, and s-win transformer, is compared using a 50% cross-validation method.
In the second part of the study, the researchers aim to predict the pathological TBSRTC class I of thyroid nodules FNAB using a CNN model based on ultrasound images. Two datasets are used for this analysis: dataset 1 includes cases from the thy-wch dataset, and dataset 2 includes cases from the Department of Ultrasound Medicine at West China Tianfu Hospital. Clinical data and ultrasound image characteristics are collected from the datasets, and univariate and multivariate analyses are performed to determine the risk factors for TBSRTC class I. The CNN model is trained using the collected data, and its performance is compared with that of clinicians in identifying TBSRTC class I nodules. Additionally, a joint diagnosis mode that provides prediction probabilities to doctors is also tested.
The results of the study show that the deep learning model for needle position recognition achieves high feasibility and accuracy. The vit base model, swin small model, and swin tiny model demonstrate similar efficiency and are suitable for different scenarios, while the resnet-50 model is not suitable for this task. The ultrasound-based prediction model for TBSRTC class I also shows high feasibility and accuracy. The accuracy rate of the CNN model is significantly higher compared to other advanced ultrasound image classification methods. The accuracy rate of TBSRTC class I predicted by clinicians with the assistance of the model is also improved to varying degrees.
In conclusion, this study demonstrates the potential of using CNN models based on ultrasound images for predicting TBSRTC class I pathology in thyroid nodules. The developed models provide an objective and efficient diagnostic tool for physicians, reducing the proportion of class I nodules and avoiding unnecessary repeat punctures.
原文地址: https://www.cveoy.top/t/topic/STS 著作权归作者所有。请勿转载和采集!