Classificação de Exames PET de Corpo Inteiro usando Representações MIP e Aprendizado Profundo

Abstract

Cancer is one of the greatest global public health challenges, with an estimated 35 million new cases by 2035. In this context, Positron Emission Tomography (PET) is essential for diagnosis and monitoring. However, the clinical interpretation of these exams is an exhaustive task, subject to the specialist’s subjectivity and limited by the high complexity of 3D volumetric data. This work proposes a method for the automatic classification of whole-body PET scans of patients with lung cancer, lymphoma, melanoma, and healthy individuals, using deep learning techniques applied to Maximum Intensity Projection (MIP) representations. The method is structured in four stages: generation of MIP images in the coronal and sagittal axes, preprocessing, feature extraction, and classification. Six architectures for feature extraction (ConvNeXt, EfficientNet-B0, Swin, and VGG19) and three classifiers (MLP, SVM, and XGBoost) were evaluated. The method achieved results of 96.45% for the AUC metric, 91.98% for the accuracy, 91.63% for the F1-Score, 91.18% for the sensitivity, and a precision of 92.08%. These results show that the use of MIP representations, combined with a set of perspective-specific specialized architectures, allows for satisfactory performance, approaching approaches that use 3D volumes and hybrid examinations (PET/CT).