Detecção de artefatos em cintilografia miocárdica utilizando inteligência artificial

Abstract

This work proposes the use of artificial intelligence techniques, specifically convo- lutional neural networks and nnU-Net, to enhance the identification of artifacts in myocardial perfusion scintigraphy images, with the goal of assisting the techni- cal team in the initial assessment of image quality for the exam in question. The methodology involved data collection in a clinical setting, followed by data prepara- tion and analysis using the 3D Slicer platform for manual segmentation of the left ventricle in the initial stage, culminating in the application and validation of the mo- dels using nnU-Net. The project then progressed to a second phase, which included manual segmentation of artifacts and a new application of AI models for automa- ted segmentation, with a focus on improving the precision of artifact identification in myocardial perfusion scintigraphy. Based on the preliminary results of the left ventricular segmentation demonstrated by internal validation, with Dice coefficient Dc = 87% and σ ± 6% and external validation of Dc = 88.46% with standard devi- ation of σ ± 7.48% and agreement of at least 80% of the segmentations performed, the process continues with the artifact segmentations. As a final result, we obtained the application of different optimization models with a Dice coefficient ranging from 0.300 to 0.940, a threshold of 0.853, the area under the ROC curve was AUC= 0.946, perfect specificity in all scenarios and modest sensitivity, indicating that the model tends not to signal some cases of interference between artifacts and the left ventri- cle. Furthermore, external validation with specialist doctors demonstrated a value of more than 60% agreement between the masks made by the specialist and the AI models, demonstrating the usefulness and assistance that the use of computational models applied in medical practice can provide to the healthcare team.