Automatic segmentation of temporal lobe epilepsy resection Cavity based on deep learning model

View at publisher Summarise with VIRAT
Authors: Lu Yang, Fengyi Yang, Zuhong Lu, Yu Sun
Cited by: 0
References: 9

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

Accurate segmentation of postoperative resection cavities on magnetic resonance imaging (MRI) is essential for the quantitative evaluation of surgical efficacy in patients undergoing anterior temporal lobe resection (ATLR) for drug-resistant temporal lobe epilepsy (TLE), yet conventional manual delineation remains subjective, labor-intensive, and poorly scalable, thereby limiting large-scale outcome analyses and the clinical translation of imaging-derived biomarkers. To address these limitations, this study proposes a three-dimensional attention-enhanced residual U-Net specifically designed to accommodate the anatomical complexity and signal heterogeneity of postoperative temporal lobe MRI. The model was trained on a large cohort of 297 ATLR cases from the publicly available IDEAS dataset, comprising high-resolution postoperative T1-weighted images with expert-validated resection cavity annotations, and the training pipeline incorporated standardized preprocessing, intensity harmonization, and spatially adaptive data augmentation to enhance robustness across diverse postoperative morphologies. Independent evaluation on a held-out test cohort of 45 subjects demonstrated strong segmentation performance, achieving a mean Dice similarity coefficient of 0.873 ± 0.041 and a mean Jaccard index of 0.781 ± 0.052, with precise boundary delineation reflected by a 95th-percentile Hausdorff distance of 3.62 ± 1.14 mm and an average surface distance of 1.42 ± 0.36 mm. Volumetric fidelity was maintained, yielding an absolute volume deviation of 6.8 ± 3.4%, thereby enabling reliable extraction of clinically meaningful resection metrics. Compared with baseline architectures, the proposed model improved Dice performance by 6.5% relative to a standard 3D U-Net and by 3.2% relative to a residual U-Net, while reducing boundary error by 1.5–2.0 mm, surpassing performance levels reported in prior studies based on smaller cohorts. In summary, the proposed framework provides a robust, scalable, and reproducible tool for automated cavity segmentation, which can facilitate large-scale surgical outcome research and be integrated into AI-assisted neurosurgical workflows.

Fields of Study & Subjects

Medical Image Segmentation Techniques Brain Tumor Detection and Classification Medical Imaging and Analysis

Publication details

Journal / source
International Conference on Communication Networks and Machine Learning (CNML)
Pages
769–773
Publication date
30 January 2026
Language
EN