In this study, we evaluated whether a machine learning classifier can automatically assess the image quality of PET images by simulating shorter acquisition times and lower quality. The main findings of our study are: First, machine learning reliably assesses subjective PET image quality. Second, its classification power is comparable to manual SNR measurements commonly used as an objective proxy for image quality. Third, the classifier performance did not differ between different reconstruction settings of her BSREM with beta values of 450 and 600. Fourth, the classifier most frequently misclassified images within image groups that experienced readers found to be on the border between non-diagnostic and diagnostic. Diagnostic value is limited. This finding further emphasizes the usefulness of robust automatic classifiers in the context of highly subjective visual assessments. Fifth, the classifier activation map matches noisy regions in non-diagnostic images, confirming the effectiveness of the non-intuitive algorithm.
Previous studies have shown that PET acquisition times can be safely shortened within certain limits without compromising image quality.19, 20, 21while others have suggested a dose regimen based on BMItwenty two Instead of current weight-based EANM recommendationstwenty three. All these suggestions are based, at least in part, on image quality ratings by experienced readers.
The utility of machine learning in PET has been demonstrated, for example, by detecting F18-FDG-PET-positive nodules and lung cancer.16,17 And in terms of image quality, we recently denoised the reconstructed F18-FDG-PET images.twenty four. However, there is a major unmet clinical need to perform large-scale image analysis to reduce both dose and scan time without compromising the modality’s diagnostic accuracy, which is a major unmet clinical need for the image quality of his PET itself. Not used for estimation. To our knowledge, our work focuses on candidate machine learning classifiers, more specifically deep neural This is the first demonstration of the capabilities of the network classifier. Compared to the radiomics approach used in a previous study involving 112 patients, the machine learning classifier achieved a higher AUC (AUC 0.978 vs 0.798 for training dataset and 0.675 for test dataset).3).
A more widespread application would be personalized clinical imaging (e.g., iterative on-the-fly evaluation of PET images to build personalized imaging protocols and adjust acquisition times until the desired image quality is achieved). ), research and the establishment of agreed standards. -Regarding image quality standards. Additionally, classifiers can be assigned to similar quality assessment tasks in PET imaging whenever a qualified subjective judgment is required.
Our study has some limitations. Its retrospective size, relatively small cohort, and use of only one scanner all limit its generalizability. Further studies are warranted to ensure broader applicability of our results. Second, the image quality assessed in our study is a strictly qualitative and subjective measure. The robustness of quantitative parameters such as standardized uptake values needs to be ensured by further studies. Third, the classifiers used in the study were not specifically validated for this task. Further optimization of the classifier could further improve the performance of image quality assessment. Fourth, reader-based ratings, the reference standard used to train this algorithm in this study, are ultimately subjective. To ensure broader applicability and acceptance of automated assessment, further studies involving a larger number of experienced readers should be conducted to establish an acceptable consensus on image quality.
