TL;DR
This paper introduces a novel 3D Active Shape Model (ASM) for automatically segmenting the eye in MRI scans, crucial for retinoblastoma treatment planning. By leveraging statistical shape information and MRI intensity, their method achieves high accuracy and significantly reduces segmentation time, offering a valuable tool for clinicians.
What Problem Does This Paper Address?
Accurate segmentation of the eye's anatomy in 3D Magnetic Resonance Imaging (MRI) is a critical but challenging task for retinoblastoma treatment planning. Retinoblastoma is a common intraocular tumor in children, and precise delineation of the eye's structures is essential for diagnosis, monitoring tumor extent, and planning radiation therapy. Existing methods often rely on manual segmentation or parametric models that require expert input and have limited adaptability to individual anatomical variations. This paper aims to bridge this gap by developing a fully automatic segmentation framework that is both accurate and efficient.
Methods
The core of this work is a 3D Active Shape Model (ASM) built upon a Point-Based Shape Variation Model (PBSVM). The process begins with constructing a baseline atlas of the eye from a dataset of healthy children's MRI scans. This atlas is then transformed into a Natural Coordinate System (NCS). A PBSVM is created by extracting surface point clouds from the atlas for different eye regions (sclera+cornea, vitreous humor, lens) and applying Principal Component Analysis (PCA) to capture shape variations.
For new patient scans, an image equalization technique is applied to standardize MRI intensity values, followed by anisotropic diffusion filtering to reduce noise and enhance anatomical edges. The ASM then combines the PBSVM with intensity information. The segmentation process involves an iterative fitting of the model to the patient's MRI, where landmark points on the model are adjusted to match the image intensity profiles. The method is designed to be fully automatic, requiring minimal user interaction.
Datasets
The study utilized a dataset of 24 healthy children's eyes (aged 3.29 ± 2.15 years) acquired using a 3T Siemens Verio MRI scanner with gadolinium-enhanced T1-weighted GE VIBE sequences. Manual segmentation of these volumes was performed by an expert radiologist to serve as the ground truth for training and validation. The method was also tested on two pathological patient eyes with retinoblastoma.
Key Insights and Ideas
- Statistical Shape Modeling for MRI: The paper demonstrates the effectiveness of Statistical Shape Models (SSM), specifically ASMs, for segmenting eye structures in MRI, a modality that presents different intensity characteristics compared to CT or Ultrasound.
- Fully Automatic Segmentation: A significant contribution is the development of a fully automatic pipeline, eliminating the need for manual landmark selection, which is a bottleneck in many existing methods.
- Robustness to Variations: The ASM is designed to accommodate shape variations and even minor eye orientation issues, making it more adaptable to individual patient anatomy.
- Speed and Efficiency: The method achieves a remarkably fast segmentation time of approximately 14 seconds per eye, which is crucial for clinical workflow.
Key Results
The proposed method achieved high segmentation accuracy on the dataset of healthy children's eyes:
- Dice Similarity Coefficient (DSC):
- Sclera and cornea: 94.90 ± 2.12%
- Vitreous humor: 94.72 ± 1.89%
- Lens: 85.16 ± 4.91%
- Mean Distance Error: 0.26 ± 0.09 mm.
- Segmentation Time: Approximately 14 seconds per eye on average.
When applied to two patients with retinoblastoma, the model also showed robust performance, achieving an average DSC of 94% for sclera+cornea, 93.98% for vitreous humor, and 92.37% for the lens, even in the presence of tumors.
Limitations and Future Work
The authors acknowledge that the lower DSC for the lens might be due to its smaller size relative to other structures, a limitation previously reported for small regions. They also identified an outlier (Sub07) with low accuracy, attributed to the small eye size of a very young patient (4 months), suggesting that a larger dataset with more younger patients could improve performance. Future work includes applying this framework to adult eyes for conditions like uveal melanoma and focusing on quantitative evaluation with larger datasets to further improve diagnostic accuracy for retinoblastoma, particularly for detecting tumor invasion into surrounding tissues.
Summary updated