2161 BrainSignsNET: A 3D CNN Framework for Robust Anatomical Landmark Detection in Brain and Spine Imaging

View at publisher Summarise with VIRAT
Authors: Siavash Shirzadeh Barough, Abhay Moghekar
Cited by: 0

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

INTRODUCTION: Accurate localization of anatomical landmarks in 3D medical imaging is crucial for several imaging tasks. Traditional atlas-based methods are often sensitive to variations in image quality, noise, and pathological changes such as atrophy, hydrocephalus, or lesions, limiting their clinical utility. Recent advances in 3D convolutional neural networks (CNNs) offer the potential for robust, automated landmark detection across diverse datasets. Moreover, such approaches can be applied to spinal CT/MRI for landmark detection and measurements (e.g., Cobb angle in scoliosis) and neurosurgical planning (e.g., targeting deep brain stimulation electrode placement). METHODS: BrainSignsNET employs a pretrained 3D ResNet-18 encoder with two decoder branches: an attention decoder generating a global heatmap and a multi-class decoder producing individual landmark heatmaps. Landmark coordinates are extracted via soft-argmax function. The training set comprised 14,272 T1-weighted scans drawn from ADNI, BLSA, and BIOCARD cohorts, which were split into 80% training and 20% internal validation datasets. An independent external validation cohort of 200 clinical scans was sourced from the Johns Hopkins Hydrocephalus Clinic. Every scan was annotated for 23 anatomical landmarks, and model performance was quantified using two complementary metrics: the mean Euclidean distance (MED) between predicted and ground-truth landmark coordinates, and the percentage of correct keypoints within a 10 mm tolerance (PCK@10). RESULTS: External validation yielded MED ranging from 3.01 to 4.13 mm, and PCK@10 ranging from 93.99 to 99.45% across 23 landmarks. CONCLUSIONS: BrainSignsNET delivers a highly accurate, generalizable model for 3D neuroanatomical landmark detection in brain MRI. Its architecture can be readily fine-tuned for spinal landmark detection and angle calculation and for precise targeting of deep brain stimulation electrodes, underscoring its broad applicability across volumetric medical imaging.

Fields of Study & Subjects

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

Publication details

Journal / source
Neurosurgery
Volume
72
Issue
Supplement_1
Pages
290–290
Publication date
26 March 2026
Language
EN