Abstract
Background and Objective: Ventricle volume is closely related to hydrocephalus, brain atrophy, Alzheimer's, Parkinson's syndrome, and other diseases. To accurately measure the volume of the ventricles for elderly patients, we use deep learning to establish a systematic and comprehensive automated ventricle segmentation framework. Methods: The study participation included 20 normal elderly people, 20 patients with cerebral atrophy, 64 patients with normal pressure hydrocephalus, and 51 patients with acquired hydrocephalus. Imaging data was obtained through picture archiving and communication systems (PACS). Ventricular structures were manually labeled using ITK software, and imaging features were extracted via machine learning. Results: This automated ventricle segmentation method applies to both CT and MRI images across different scan slice thicknesses, producing excellent segmentation results (Dice > 0.9). Conclusion: This method has wide applicability and clinical practicability, assisting clinicians in early disease detection, diagnosis, and treatment evaluation.
Introduction
Ventricle volume serves as a critical biomarker for various neurodegenerative conditions, particularly in elderly populations where conditions like Alzheimer’s, Parkinson’s, and normal pressure hydrocephalus (NPH) are prevalent. Precise measurement of ventricular enlargement is essential for accurate diagnosis and monitoring disease progression.
Despite the importance of these measurements, automated segmentation faces significant hurdles in clinical practice. Most existing models are specialized for a single modality (either MRI or CT) and are typically trained on high-resolution "thin-slice" images (<3mm). However, clinical reality often necessitates "thick-slice" scans (>3mm) due to time and radiation constraints. Models trained on one resolution often fail on the other. Furthermore, manual segmentation—the current gold standard—is labor-intensive, taking approximately 30 minutes per patient, and is subject to inter-observer variability. This study introduces a "thickness-agnostic" deep learning framework designed to bridge these gaps.
Methods and Architecture
The researchers developed a semi-supervised, encoder-decoder architecture designed to handle the variability in clinical imaging.
The Architecture
The framework utilizes ResNet-34, pre-trained on ImageNet, as the encoder to extract robust features from the input images. For the decoder, the team implemented sub-pixel convolution instead of standard deconvolution or interpolation. This choice allows for the reconstruction of high-resolution segmentation masks without the heavy computational overhead or the "checkerboard artifacts" common in other upsampling methods.
Semi-Supervised Learning and Loss Function
A core innovation of this study is its semi-supervised approach. Because annotating thin-slice images is exceptionally difficult for clinicians, the model was trained using labeled thick-slice images (source domain) and unlabeled thin-slice images (target domain). To address the "domain shift" between these resolutions, the researchers employed a Pearson $\chi^2$ divergence loss function. Unlike standard KL-divergence, Pearson $\chi^2$ divergence effectively balances gradients between "easy" and "hard" samples, preventing the model from becoming biased toward easily segmented regions and ensuring better delineation of complex structures like the lateral ventricle horns.
Dataset and Clinical Cohort
The retrospective study utilized data from 155 participants at Shenzhen Second People’s Hospital (2016–2019). The cohort was diverse, including:
- 20 normal elderly (NE) individuals
- 20 patients with brain atrophy (BA)
- 64 patients with idiopathic normal pressure hydrocephalus (iNPH)
- 51 patients with acquired hydrocephalus (AH)
The dataset comprised over 3,600 thick-slice and 4,200 thin-slice images obtained from multiple CT and MRI (1.5T and 3.0T) scanners, providing a robust, real-world testing environment.
Key Results
The proposed framework significantly outperformed standard benchmarks, including U-Net and U-Net++.
- High Accuracy: The model achieved a Dice coefficient > 0.9 across nearly all categories. Notably, for thin-slice MRI, the model reached a Dice score of 0.9056, a substantial improvement over U-Net’s 0.7665.
- Modality and Pathology Robustness: The system demonstrated high stability on CT scans, which often suffer from lower contrast compared to MRI. It also successfully handled extreme ventricular distortions caused by tumors or trauma—cases where traditional atlas-based methods frequently fail.
Discussion and Future Directions
While the results are promising, the authors note certain limitations. The model was developed specifically for elderly patients; its performance in pediatric cases, where brain structures differ significantly, is yet to be determined. Additionally, as the data originated from a single center, further multi-center validation is required to ensure generalizability across different hospital systems.
Future work will focus on integrating biological features, such as laboratory test results, with imaging data. This holistic approach aims to create a diagnostic tool capable of not only measuring ventricular abnormalities but also predicting their underlying causes.
Conclusion
This automated ventricle segmentation method offers wide applicability and clinical practicability. By providing rapid, accurate, and modality-independent measurements, it serves as a powerful tool for clinicians to detect early-stage disease, monitor progression, and evaluate the efficacy of treatments in elderly patients.
Summary updated