Texture-Aware 3D Deep Learning in Segmenting Skull-Conjoined Brain Tumor

Tumour segmentation of brain, magnetic resonance imaging (MRI); 3-dimensional deep learning; convolutional neural networks; vision transformers; a wavelet transform; medical image analysis; brain tumor detection.

Pooja P P1,
Research scholar (Part Time), Department of
Computer Science and engineering, School of
Engineering and Technology, CHRIST
(Deemed to be University), Bangalore, and
Assistant Professor, PSAIAC, Presidency
University, Bangalore, India.
pooja.pp@res.christuniversity.in,pooja.p@presi
dencyuniversity.in

Dr Aruna S K2,
Associate Professor, Department of AI and data
Science engineering, School of Engineering
and Technology, CHRIST (Deemed to be
University), Bangalore, India,
aruna.sk@christuniversity.in

Received:10-06-2026
Revised:16-07-2026
Accepted:20-07-2026

European Journal of Prosthodontics and Restorative Dentistry (2026) 35(5s), 83–95

Texture-Aware 3D Deep Learning in Segmenting Skull Conjoined Brain Tumor

Magnetic resonance imaging (MRI) cannot provide sufficient contrast, heterogeneity of the tumor texture, and vague tumor-bone interfaces that makes it difficult and challenging to carefully segment brain tumours that are located close to the cranial skull boundary of the brain. Traditional convolutional neural network (CNN) architecture tends to lose delicate structural aspects in these areas. The paper introduces a texture-sensitive hybrid deep learning architecture to skin skull-conjoined brain tumors segmentation of multi-modal brain tumor volumes using MRI volumes. The framework combines the contrast-adaptive preprocessing, three-dimensional multi-phase learning (3D-MPTL) on the texture and three-dimensional double density dual-tree Complex wavelet transform (3D-DDDTCWT) to obtain the volumetric texture and the multi-resolution directional features. A Swin Transformer encoder is used to extract long range spatial statistics and 3D U-Net network is modified to voxel tumor segmentation. Alongside, a hybrid particle swarm optimization and gravitational search algorithm (PSOGSA) is applied to optimal network hyperparameter and training convergence. The experimental assessment of the BraTS dataset shows a better finalization of segmentation especially in the region of a skull and adjacent tumor with increased Dice scores in comparison to the current CNN and transformer-based assessment. Experimental evaluation on the BraTS dataset demonstrates that the proposed method achieves Dice scores of 97.41%, 95.62%, and 93.78% for Whole Tumor, Tumor Core, and Enhancing Tumor regions, respectively. The model also achieves a Hausdorff Distance as low as 2.15 mm, indicating accurate boundary localization. Comparative analysis shows that the proposed framework outperforms state-of-the-art models by 2–4% improvement in Dice score, particularly in challenging skull-adjacent tumor regions.

•••••••••••••••••••••••••••••••• ejprd.org - Published by Riset Publication Services LLC

EJPRD

Copyright ©2026 by Riset Publication Services LLC