TL;DR
This paper investigates the accuracy and reliability of manual 3D pelvic landmark labeling for pelvimetry and introduces an automatic labeling method. The automatic approach demonstrates comparable or superior performance to manual labeling, offering a promising avenue to reduce radiologist workload, standardize measurements, and enhance the detailed description of the pelvic cavity.
What Problem Does This Paper Address?
Cephalopelvic disproportion (CPD), a critical condition where the fetal head is too large for the maternal pelvic cavity, can lead to difficult deliveries and complications. Accurately assessing pelvic cavity morphology is crucial for predicting and managing CPD. While various imaging techniques exist, quantifying pelvic morphology often relies on identifying anatomical landmarks. However, manual labeling of these landmarks on 3D models can be time-consuming and prone to variability. This study addresses the need to evaluate the accuracy and reliability of manual 3D pelvimetry and to develop and assess an automated method that could potentially overcome the limitations of manual approaches.
Methods
The study employed a two-pronged approach:
- Manual Labeling and Pelvimetry:
- Ten anonymized pelvic CT scans were segmented into sacrum, left hip, and right hip by three operators, resulting in 30 3D voxel-based models.
- Three radiologists manually labeled 12 anatomical landmarks on these 3D pelvic models using a custom graphical interface.
- Based on these labeled landmarks, 15 pelvimetry measures (13 lengths and 2 angles) were calculated.
- The accuracy of manual labeling was assessed by comparing each radiologist's landmark placement to the average of the other two. Pelvimetry measure errors were similarly quantified.
- Inter-radiologist reliability for pelvimetry measures was evaluated using the intraclass correlation coefficient (ICC) and standard error of measurement (SEM).
- Automatic Labeling:
- A reference pelvic model was created by averaging the 30 manually labeled pelvic models and their corresponding landmarks.
- A non-rigid registration algorithm was used to match this reference model to individual target pelvic models.
- Anatomical landmarks were then projected from the matched reference model onto the target pelvis.
- The accuracy and reliability of the automatic labeling method were assessed by comparing its results to the average of the three radiologists' manual labels.
Datasets
The study utilized a dataset of 10 anonymized pelvic CT scans. These scans were segmented by three operators, resulting in a total of 30 individual 3D pelvic models for evaluation. No new benchmark datasets were introduced.
Key Insights and Ideas
- Landmark Labeling Variability: There's a significant difference in how consistently radiologists label different anatomical landmarks. Some landmarks are consistently placed with minimal error, while others are prone to larger deviations, highlighting the need for standardization.
- Limited Error Propagation: Despite variability in landmark labeling, the errors did not significantly propagate to the final pelvimetry measures. This suggests that the calculated measures are relatively robust to minor labeling inaccuracies.
- Automatic Labeling Advantage: The developed automatic labeling method not only matched but often surpassed the accuracy and reliability of manual labeling by radiologists. This is a crucial finding for future clinical applications.
- Potential for Efficiency and Standardization: Automating the labeling process could drastically reduce the time radiologists spend on pelvimetry, allowing them to analyze more cases or focus on more complex aspects of patient care. It also offers a path to standardize measurements across different institutions and operators.
- Future of CPD Prediction: The ability to generate large, standardized datasets of pelvic morphology through automatic labeling could be pivotal in improving our understanding of CPD and developing more effective predictive indices.
Key Results
- Landmark Labeling Accuracy:
- Median labeling errors for manual labeling ranged from 2.3 to 11.6 mm.
- Median labeling errors for automatic labeling ranged from 1.9 to 6.3 mm.
- For six landmarks, automatic labeling showed statistically significant improvements in accuracy compared to manual labeling.
- Pelvimetry Measures Accuracy:
- Median errors for length measures were 0.7-5.5 mm for manual labeling and 0.6-4.1 mm for automatic labeling.
- Median errors for angle measures were 2.5° (pectineal) and 4.1° (subpubic) for manual labeling, and 1.8° and 4.0° respectively for automatic labeling.
- Only one out of 15 pelvimetry measures reported a median error above 5 mm or 5° with manual labeling.
- Inter-Radiologist Reliability:
- Pelvimetry measures showed moderate to excellent inter-radiologist reliability, with 10 measures achieving excellent reliability (ICC ≥ 0.9).
- Reliability Across Segmentations:
- Manual labeling showed good to excellent reliability across segmentations.
- Automatic labeling consistently achieved excellent reliability across segmentations, outperforming manual labeling in this regard.
Limitations and Future Work
- Limited Sample Size: The study used only 10 CT scans, which, while segmented 30 times, represents a relatively small dataset. The authors acknowledge that results might not be generalizable to all radiologists or operators, and comparisons between institutions could yield different outcomes.
- Radiologist Expertise: The study involved three radiologists from a single institution. Future work could explore comparisons with radiologists from diverse backgrounds and institutions.
- Focus on Obstetric Pelvimetry: While the methods could be applied to other fields, the current study's interpretation and error thresholds are specific to obstetrics.
- Future Directions:
- Expanding Datasets: The authors emphasize the need for larger datasets to improve understanding of pelvic morphology and CPD.
- Integrating Automatic Segmentation: Combining the proposed automatic labeling with existing automatic segmentation methods could lead to a fully automated pelvimetry pipeline.
- Exploring Ultrasound: Investigating the potential of ultrasound imaging for 3D pelvic imaging, given its widespread use in obstetrics, is a promising avenue.
- Fetal Head Analysis: Extending automatic labeling to the fetal head could provide a more comprehensive assessment of CPD.
- Clinical Integration: Further studies are needed to integrate morphological and clinical data to improve the clinical utility of pelvimetry measures.
Summary updated