TL;DR
This paper introduces a novel method for automatically segmenting the prostate in transrectal ultrasound (TRUS) videos. By combining global prostate shape knowledge with patient-specific, locally adaptive shape statistics, the algorithm significantly improves segmentation accuracy, particularly in challenging base and apex regions, paving the way for more precise image-guided prostate interventions.
What Problem Does This Paper Address?
Accurate segmentation of the prostate in transrectal ultrasound (TRUS) images is crucial for guiding prostate biopsies and therapies, and for fusing TRUS data with other imaging modalities like MRI. However, segmenting the prostate in TRUS is notoriously difficult due to significant variations in prostate shape, especially at the base and apex, coupled with low signal-to-noise ratios in the ultrasound images. Existing methods often struggle with these challenging regions, leading to inaccurate boundaries.
Methods
The proposed method employs a deformable contour-based segmentation approach. The core innovation lies in its use of a hybrid shape constraint that combines:
- Global Population-based Shape Statistics (GPSS): This provides a general understanding of prostate shapes learned from a large dataset of manually segmented prostates.
- Adaptive Local Shape Statistics (ALSS): This is the key novelty. Instead of relying solely on global statistics, ALSS learns patient-specific shape characteristics from a sliding window of recently segmented frames. This allows the model to adapt to the unique shape of a particular patient and, crucially, to capture rapid, local shape changes as the TRUS probe moves through different regions of the prostate (base, mid-gland, apex).
The segmentation process is initialized using GPSS for the first few frames. As more frames are segmented, ALSS is computed and then used to re-segment previous frames and guide the segmentation of subsequent frames. This adaptive updating mechanism ensures that the shape prior is always relevant to the current region being segmented. The energy function for the deformable contour includes image features, internal contour constraints (smoothness), and the shape constraint (GPSS or ALSS). Optimization is performed using a two-step process involving an iterative shape fitting method and dynamic programming for efficient minimization.
Datasets
The study was validated using 19 TRUS video sequences, totaling 3064 frames, from 19 different patients undergoing prostate cancer biopsy. Manual segmentations performed by an experienced radiologist served as the ground truth. For training the global shape statistics (GPSS), a number of manually segmented contours from different subjects' TRUS videos were used. For ALSS, the segmented shapes from the current video sequence itself were used.
Key Insights and Ideas
- Local Adaptation is Key: The most significant insight is that relying solely on global shape statistics is insufficient for accurately segmenting the prostate, especially in the highly variable base and apex regions. Adapting shape statistics locally to the patient and the specific region being scanned is critical.
- Dynamic Shape Priors: The paper demonstrates the power of using a "living" shape prior (ALSS) that continuously updates based on newly acquired segmentation data. This allows the model to track subtle but important shape changes.
- Improved Base and Apex Segmentation: The method's ability to capture local shape variations directly translates to substantially better performance in the challenging base and apex regions, which are often problematic for other methods.
- Robust Initialization: The two-stage initialization process (GPSS followed by ALSS re-segmentation) makes the method robust to initial segmentation errors.
Key Results
- Overall Accuracy: The proposed method achieved an average Mean Absolute Distance (MAD) error of 1.65 ± 0.47 mm. This is a significant improvement over independent frame segmentation (2.50 ± 0.81 mm) and frame propagation (2.01 ± 0.63 mm).
- Base and Apex Improvement: The method reduced MAD errors by 49.4% and 18.9% in the base (compared to Method-1 and Method-2 respectively) and by 55.6% and 17.7% in the apex.
- Speed: The algorithm segments each frame in approximately 200 ms on a Core2 1.86 GHz PC, making it suitable for real-time applications.
- Qualitative Results: Visual results show that the proposed method (ALSS-6) produces segmentations that closely follow the prostate boundaries, particularly in the base and apex, where other methods show noticeable deviations.
Limitations and Future Work
- Limitations: While the method significantly improves base and apex segmentation, the paper notes that the performance of ALSS can be affected by the choice of the number of training shapes (N). If N is too small, the segmentation can become too restrictive; if N is too large, ALSS may not adapt quickly enough. The optimal N was found to be 6.
- Future Work: The authors suggest that this adaptive local shape statistics approach could be extended to 3D TRUS video segmentation in the future. They also highlight that the method's robustness and accuracy in challenging regions make it a promising tool for improving image-guided prostate interventions.
Summary updated