%0 Journal Article
%T 基于深度学习的无参考CT图像质量评估模型
A No-Reference CT Image Quality Assessment Model Based on Deep Learning
%A 沈菁莹
%A 章浩伟
%A 刘颖
%J Modeling and Simulation
%P 186-198
%@ 2324-870X
%D 2025
%I Hans Publishing
%R 10.12677/mos.2025.143214
%X 无参考CT图像质量评估(NR-IQA)旨在建立与放射科医生主观评估高度一致的客观图像质量评估体系。目前诸多临床CT图像数据集没有实际IQA评分,基于此,本研究提出一种基于深度学习的NR-IQA模型,并对其进行验证。NR-IQA模型将卷积模块(CNN)与视觉Transformer模块(ViT)结合,同时训练4个CNN-ViT网络集成作为教师模型,以模拟放射科医生多次主观IQA过程;接着结合知识蒸馏框架,将教师模型的信息蒸馏到1个学生模型(单一CNN-ViT网络)中。本研究结合峰值信噪比(PSNR)和结构相似性(SSIM)两个客观指标来评估CT图像质量,并使用其标注临床胸部CT图像数据集以验证NR-IQA模型。提出的NR-IQA模型综合性能达到2.8070,PLCC为0.9916,SROCC为0.9683,KRCC为0.8471,MAE低至0.0259,MSE仅为0.0010,验证了其预测CT图像IQA精度的优越性。
No-reference CT image quality assessment (NR-IQA) aims to establish an objective image quality evaluation system that achieves high consistency with radiologists’ subjective assessments. Given the current lack of actual IQA scores in numerous clinical CT image datasets, this paper proposed and validated a NR-IQA model based on deep learning. The proposed model integrated convolutional neural network (CNN) modules with visual Transformer (ViT) modules, and trained an ensemble of four CNN-ViT networks as teacher models to simulate radiologists’ repeated subjective IQA processes. Subsequently, a knowledge distillation framework was employed to transfer the information from teacher models to a student model (a single CNN-ViT network). This paper combined two objective metrics, peak signal-to-noise ratio (PSNR) and structural similarity (SSIM), to evaluate the quality of CT images, and used them to annotate the CT image dataset to validate the proposed NR-IQA model. The comprehensive performance of the proposed NR-IQA model reached 2.8070, achieving a Pearson linear correlation coefficient (PLCC) of 0.9916, Spearman rank-order correlation coefficient (SROCC) of 0.9683, Kendall rank correlation coefficient (KRCC) of 0.8471, with mean absolute error (MAE) reduced to 0.0259 and mean squared error (MSE) as low as 0.0010, validating its superior accuracy in predicting CT image IQA scores.
%K 无参考图像质量评估,
%K 临床胸部CT图像,
%K 视觉Transformer,
%K 知识蒸馏
NR-IQA
%K Clinical Chest CT Images
%K Vision Transformer
%K Knowledge Distillation
%U http://www.hanspub.org/journal/PaperInformation.aspx?PaperID=109387