Akademik Veri Yönetim Sistemi
Biomedical Signal Processing and Control, cilt.112, 2026 (SCI-Expanded, Scopus)
There is no other disease that affects life expectancy as seriously as brain tumors. Their accurate and rapid segmentations from images are vital for treatment planning and surgery operations. Therefore, in this work, a narrowing pyramid model was developed using transformer and convolutional network pairs to provide efficiency in the detection and extraction of tumor features. Also, a composite loss function was designed by utilizing shape, probability distribution, and border information to improve the feature discrimination ability. Moreover, a multi-head merge-attention mechanism was constructed and integrated into the transformer structure to establish strong links between information sequences. Extensive experiments were conducted with images from two public databases, the Brain Tumor Segmentation 2023 and 2024 challenge databases. In addition, state-of-the-art approaches were applied using the same datasets and assessed using identical criteria for equitable comparative analyses. Therefore, this study's major contributions are: (i) introducing a network designed as a narrowing pyramid with transformer and convolutional network pairs; (ii) introducing a composite loss function that takes into account probability distributions, shapes, and border information; (iii) introducing a merge-attention mechanism that creates strong links between information sequences; (iv) presenting equitable comparisons of state-of-the-art methods applied to identical datasets and evaluated with identical criteria; (v) presenting thorough experimental results, demonstrating the proposed network's effectiveness in terms of average lesion-wise dice score (97.49 %) and lesion-wise Hausdorff distance (2.40 mm) over recent methods.