A deep learning AI model for determining the relationship between X-Ray detectors and patient positioning in chest radiography

Abstract

PURPOSE: The objective of this study was to create an artificial intelligence (AI) system capable of automatically detecting the positional relationship between an X-ray detector and the patient during anteroposterior chest radiography. METHODS: In this study, a total of 22299 images depicting the positional states of X-ray detectors relative to patients were used to develop an AI system for the automatic determination of X-ray detector status. The images were captured from Routine clinical chest X-ray radiography practice settings, without exposing any patient privacy, adhering to the Declaration of Helsinki. The PyTorch library was utilized for customizing a Convolutional Neural Network (CNN) model for the training of the chest radiography positional determination model. RESULTS: The average accuracy of Model A on the validation set was 0.9668, with an average loss function value of 0.1078. In contrast, Model B achieved an average accuracy of 0.9776, also with an average loss function value of 0.0970. In the test set results, Model A -fold 4 demonstrated a true negative rate (TNR) of 0.9925, negative predictive value (NPV) of 0.9925, precision of 0.9699, recall of 0.9700, accuracy of 0.9700, and an F1 score of 0.9699. Model B -fold 1exhibited a TNR of 0.9946, NPV of 0.9946, precision of 0.9787, recall of 0.9784, accuracy of 0.9784, and an F1 score of 0.9784. McNemar's test indicated statistical difference between the two models. CONCLUSION: The AI model utilizing a customized CNN architecture has demonstrated its potential to automatically detect the positional relationship between the patient and the X-ray detector during chest radiography procedures. This model can potentially alleviate the workload of radiologic technologists in producing chest radiographs and enhance the accuracy of the imaging process.

Ähnliche Arbeiten