基于深度学习的胸部X线图像清晰度评价方法
Method Based on Deep Learning for Evaluating Clarity of Chest X-ray Images
宋亮亮 1王倩 1韩啸 2李传富 3李小虎 4余永强4
作者信息
- 1. 安徽医科大学第一附属医院影像科,安徽 合肥 230022
- 2. 安徽影联云享医疗科技有限公司,安徽 合肥 230088
- 3. 合肥综合性国家科学中心先进研究院(安徽省人工智能实验室),安徽 合肥 230088;安徽中医药大学第一附属医院影像科,安徽 合肥 230012
- 4. 安徽医科大学第一附属医院影像科,安徽 合肥 230022;安徽省影像诊断质控中心,安徽 合肥 234099
- 折叠
摘要
目的 构建深度学习模型对胸部X线图像进行清晰度评价,并与放射科医师的主观评价对比,验证模型的效能.资料与方法 回顾性收集 2015 年 6 月—2022 年 8 月安徽省 590 家医院共 9 135 幅胸部X线图像,组织放射科医师采用五级评分法对图像清晰度进行多人多次评价,单人评价结果为A、B,多人评价结果为C.构建基于ResNet-50的深度学习模型对胸部X线图像进行清晰度评价,以结果C作为模型训练和测试数据,模型评价结果为D.由 1名放射质控专家对模型评价结果和医师多人评价结果进行审核评价作为图像清晰度的参考标准,评价结果为E.采用Spearman相关、均方根误差(RMSE)和准确率验证模型的效能.结果 与参考标准E相比,D的平均准确率为0.85,高于C的0.84.A、B、C、D与E的ρ分别为0.58(0.54,0.62)、0.59(0.55,0.63)、0.74(0.71,0.77)和0.80(0.78,0.82),D与E的相关性最好.A与B的ρ为0.45(0.41,0.49),两次单人主观评价清晰度相关性较差.A、B、C、D与E的RMSE分别为0.99、0.94、0.72和0.71,D与E的RMSE小于人工评价结果.结论 本研究构建的模型能够准确评价胸部X线图像清晰度,通过深度学习方法可以降低人工评价的主观干扰,为临床放射图像清晰度评价提供有效、客观的工具.
Abstract
Purpose Develop deep learning models to assess the clarity of chest X-ray images and validate the model's effectiveness by comparing it with the subjective evaluations of radiologists.Materials and Methods A retrospective collection of 9 135 chest X-ray images from 590 hospitals in Anhui Province,spanning from June 2015 to August 2022,was organized involving multiple radiologists who repeatedly evaluated the clarity of the images using a five-level scoring system.Individual assessments were designated as A and B,whereas the collective result of multiple assessments was designated as C.By constructing a deep learning model based on ResNet-50,image clarity evaluations of chest X-ray images were performed,we used the result C as the training and testing data for the model.The model's evaluation results were denoted as D.A radiology quality control expert conducted an audit assessment of the model's evaluation results and the multi-person assessments of physicians,serving as the reference standard for image clarity.Their assessment results were labeled as E.Statistical analysis,including Spearman's rank correlation coefficient,root mean square error(RMSE)and accuracy was employed to evaluate the effectiveness of the model.Results Compared with the reference standard E,D achieved an average accuracy of 0.85,exceeding the accuracy of C,which stood at 0.84.The ρ values for A,B,C,D and E were 0.58(0.54,0.62),0.59(0.55,0.63),0.74(0.71,0.77)and 0.80(0.78,0.82),respectively.The model exhibited the highest correlation with E.The ρ between A and B was 0.45(0.41,0.49),indicating a lower correlation between two individual subjective assessments of image clarity.The RMSE values for A,B,C,D and E were 0.99,0.94,0.72,and 0.71,respectively.The model's RMSE was lower than that of manual assessments.Conclusion This research model is capable of accurately assessing the clarity of chest X-ray images,and reducing the subjective interference of manual evaluation through deep learning methods,thereby providing an effective and objective evaluation tool for the assessment of clarity in clinical radiographic images.
关键词
深度学习/质量控制/放射摄影术,胸部/决策,计算机辅助Key words
Deep learning/Quality control/Radiography,thoracic/Decision making,computer-assisted引用本文复制引用
基金项目
安徽省高等学校协同创新项目(GXXT-2021-065)
出版年
2024
NETL
NSTL
万方数据