Prediction of TP53 mutations across female reproductive system pan-cancers using deep multimodal PET/CT radiogenomics.

Background: TP53 mutations play a critical role in the clinical management and prognostic evaluation of gynecologic malignancies such as cervical, endometrial, and ovarian cancers. With the advancement of radiomics and deep learning technologies, noninvasive AI models based on medical imaging have become important tools for assessing TP53 mutation status. Methods: This study retrospectively analyzed 259 patients with cervical, endometrial, or ovarian cancer who underwent PET/CT before treatment. Radiomics features from tumors and brown adipose tissue (BAT) were extracted, and a Transformer-based model was developed to predict TP53 mutation by integrating imaging and clinical data. The model was trained with five-fold cross-validation, and clustering analysis was performed on deep features to explore their correlation with TP53 status. Results: Radiomic features from tumor CT images, tumor PET images, brown adipose tissue CT images, and brown adipose tissue PET images were all found to be associated with TP53 mutation status in gynecological tumors. On the test set, the accuracy of the tumor CT radiomic model was 0.7931, the tumor PET radiomic model achieved an accuracy of 0.8276, the brown adipose tissue CT radiomic model had an accuracy of 0.7241, and the brown adipose tissue PET radiomic model reached an accuracy of 0.7931. The combined model achieved an accuracy of 0.8620 on the test set, and after automatic annotation using nn-UNet, the combined model's accuracy was 0.8000. Unsupervised clustering of the deep features extracted by the combined model showed that the image clustering patterns were significantly correlated with TP53 mutation status ( Conclusion: This study demonstrates that radiomic features from tumor and brown adipose tissue CT and PET images are closely associated with TP53 mutation status in gynecological tumors. This study constructed a cross-cancer TP53 model. The combined model constructed based on multi-modal imaging effectively captures TP53-related imaging phenotypes across different cancer types, and these phenotypic patterns show a significant correlation with TP53 mutation status.