TY  - JOUR
T1  - Modality-Efficient Deep Learning for Automated Multi-Class Segmentation of High-Grade Gliomas
A1  - Gabriel Costa
A1  - Lucas Ribeiro
A1  - Ricardo Alves
A1  - Mariana Lopes
JF  - Journal of Medical Sciences and Interdisciplinary Research
JO  - J Med Sci Interdiscip Res
SN  - 3108-4826
Y1  - 2025
VL  - 5
IS  - 2
DO  - 10.51847/RjebAFYcFF
SP  - 188
EP  - 206
N2  - Glioma constitutes the most frequent primary malignant neoplasm of the brain. Precise delineation of glioma regions on magnetic resonance imaging (MRI) is vital for clinical treatment planning. Conventional approaches to glioma segmentation typically depend on four MRI sequences: T1, T2, T1Gd, and FLAIR. Obtaining all four sequences prolongs patient exposure to the scanner and extends the overall computational time required for segmentation. In practice, however, complete sets of these modalities are not always available because of scanner time limitations or patient-related challenges. To address this issue, the present study applied U-Net-based fully convolutional neural networks for automatic segmentation, specifically investigating whether a reduced set of MRI modalities significantly impairs segmentation performance. The developed framework was trained, validated, and tested on 100 high-grade glioma (HGG) cases under two conditions: first with the complete set of MRI sequences and second with only FLAIR and T1Gd. Evaluation on the held-out test data revealed that the baseline U-Net achieved a mean Dice score of 0.9166 across all sequences and 0.9190 when restricted to FLAIR and T1Gd. To improve performance using only these two modalities, we adopted an ensemble strategy, incorporating pre-trained VGG16, VGG19, and ResNet50 architectures as modified encoders within the U-Net structure for glioma segmentation. We trained, validated, and tested these models on a larger cohort of 259 high-grade glioma (HGG) cases. The experimental outcomes demonstrated mean Dice scores of 0.9395 for the baseline U-Net and 0.9360, 0.9359, and 0.9356 for the ensembles based on VGG16, VGG19, and ResNet50 encoders, respectively. We also compared performance with prior research using the full complement of four MRI modalities. This investigation underscores that FLAIR and T1Gd contribute most to accurate tumor segmentation. The baseline U-Net model proved robust in segmenting sub-tumoral components of HGG and exhibited results comparable to current state-of-the-art methods.
UR  - https://smerpub.com/article/modality-efficient-deep-learning-for-automated-multi-class-segmentation-of-high-grade-gliomas-xz1tgnhutyhjey7
ER  -