Synthetic rT1/T2 maps as a novel input modality for backbone U-Net glioma sub-region segmentation

Accurate glioma sub-region segmentation is critical for treatment monitoring and surgical planning. Recently, researchers have shown an increased interest in using deep neural networks (DNN), such as backbone U-Nets, for glioma segmentation. Input data is a fundamental property of DNN models training. This study investigated reconstructed relative T 1 /T 2 (rT 1 /T 2 ) maps as an input modality for backbone U-Net glioma segmentation. In this study, 381 cases from the BraTS 2023 dataset were included. rT 1 /T 2 maps were reconstructed from preprocessed T 1 -weighted and T 2 -weighted sequences. Backbone U-Nets with six backbone architectures, ResNet34, ResNet152, InceptionV3, VGG16, DenseNet121, and MobileNet, were single-channel trained using 5-fold cross-validation and patient-wise train-test splitting. Performance was evaluated using Dice Similarity Coefficient, Intersection over Union, and Hausdorff Distance across peritumoral edema, necrotic core, and active tumor regions. TOPSIS multi-criteria analysis was applied for objective backbone-modality ranking. ResNet34 backbone U-Net trained with rT 1 /T 2 maps was optimal for edema (DSC = 0.901, HD95 = 1.57 mm, and CC = 0.997) and necrotic core (DSC = 0.816, HD95 = 2.43 mm, and CC = 0.994), while InceptionV3 backbone U-Net trained with rT 1 /T 2 maps demonstrated superior active tumor segmentation (DSC = 0.837, HD95 = 1.66 mm, and CC = 0.975). In the TOPSIS ranking, rT 1 /T 2 ranked first overall with a CC of 0.794, and InceptionV3 was the top-ranked backbone with a CC of 0.938. Within this single-channel experimental setting, reconstructed rT 1 /T 2 maps showed a modest but consistent average advantage over conventional T 1 -weighted and T 2 -weighted sequences across backbone architectures for backbone U-Net-based glioma segmentation. Whether this advantage reflects the ratio transformation specifically, or the fact that rT 1 /T 2 draws on information from both sequences, was not tested here and warrants further investigation, including comparison against a combined multi-channel T 1 + T 2 input.