Differential Diagnosis of Ovarian Tumor Using FDG PET/CT, based on Convolutional Neural Networks.
Authors
Affiliations (1)
Affiliations (1)
- Department of Nuclear Medicine, Busan Paik Hospital, University of Inje College of Medicine, 75, Bokji-ro, Busanjin-gu, Busan, Republic of Korea.
Abstract
This study aims to evaluate the utility of convolutional neural networks (CNNs) in the differential diagnosis of benign and malignant ovarian tumors using FDG PET/CT imaging. A total of 101 patients with pelvic masses who underwent F-18 FDG PET/CT before surgery between January 2020 and December 2023 were included. Patients were classified into benign and malignant tumor groups based on histopathological diagnosis. FDG PET/CT images were analyzed using two CNN models: a ResNet-18 model and a simple CNN architecture. The models were evaluated using accuracy, area under the curve (AUC), sensitivity, and specificity, with Gradient-weighted Class Activation Mapping (Grad-CAM) applied to visualize important regions within the images. The ResNet-18 model achieved an accuracy of 0.882, an AUC of 0.938, sensitivity of 0.829, and specificity of 0.913. The simpler CNN model achieved an accuracy of 0.876, an AUC of 0.957, sensitivity of 0.761, and specificity of 0.942. Grad-CAM heatmaps identified regions of the PET/CT images that the models found most relevant for classification. This study demonstrates the potential of deep learning-based FDG PET/CT analysis for differentiating benign and malignant ovarian tumors. The CNN model showed high diagnostic accuracy, emphasizing the value of CNN-based models in PET-based tumor classification.