TL;DR
This paper introduces a deep learning-assisted CT volumetry method for estimating the weight of liver grafts in living donors. By automating liver segmentation, the approach significantly speeds up the process and provides an accurate conversion formula, crucial for ensuring donor safety and recipient outcomes in liver transplantation.
What Problem Does This Paper Address?
Living-donor liver transplantation (LDLT) is a life-saving procedure, but it hinges on having an adequate graft size. Too small a graft can lead to recipient complications, while harvesting too much can endanger the donor. Traditionally, CT volumetry (CTV) is used to estimate the graft weight (GW) before surgery. However, this process is often time-consuming due to manual liver segmentation. Furthermore, previous studies have used small sample sizes or questionable assumptions to derive volume-to-weight conversion formulas, leading to potentially inaccurate estimations. This paper aims to address the inefficiency and accuracy limitations of current CTV methods for GW estimation in LDLT.
Methods
The core of this study is the application of a deep learning algorithm (DLA) for automated liver segmentation on CT scans. This DLA was integrated into software that first automatically segmented the liver. Radiologists then reviewed and corrected any segmentation errors, a process that was found to be minor and quick. The software then calculated the graft volume based on a defined resection plane.
A retrospective study was conducted on 581 living liver donors who underwent right-lobe graft donation. The data was split into two groups: a development group (n=207) used to derive a volume-to-weight conversion formula via linear regression, and a validation group (n=374) to test the formula's accuracy. The key innovation here is the use of DLA for automated segmentation, which drastically reduces the manual effort and time required for CTV.
Datasets
The study utilized retrospective CT imaging data from 581 living liver donors who underwent right-lobe graft donation at a single institution between 2013 and 2015. This dataset was divided into a development group (donations in 2013) and a validation group (donations from 2014-2015). A subset of 50 donors from the validation group was used for inter-reader agreement assessment. No new benchmark datasets were introduced.
Key Insights and Ideas
- Deep Learning for Efficiency: The DLA significantly streamlines the CTV process. Automating liver segmentation means radiologists spend less time on tedious manual outlining, freeing them up for more critical tasks.
- Accuracy Matters for Safety: In LDLT, precise graft weight estimation is not just about surgical planning; it's directly linked to donor safety and recipient outcomes. Even small errors can have significant consequences.
- Large Cohort for Robust Formulas: Developing a conversion formula from a large dataset (207 donors) in the development group is crucial for its reliability and generalizability compared to formulas derived from very small sample sizes.
- Beyond Simple Correlation: The study goes beyond just looking at correlations and focuses on the measurement error using Bland-Altman analysis, which is a more clinically relevant metric for understanding the range of potential discrepancies.
Key Results
- Time Efficiency: The DLA-assisted CT volumetry process, including review and correction, took an average of 1.8 ± 0.6 minutes, a significant improvement over traditional manual methods.
- Conversion Formula: A robust volume-to-weight conversion formula was developed: estimated graft weight (g) = 206.3 + 0.653 × CT-measured graft volume (mL).
- Accuracy in Validation: In the validation group, the Bland-Altman 95% limit-of-agreement (LOA) for estimated vs. measured graft weight was -1.7% ± 17.1%. This indicates a mean bias of -1.7% and a measurement error of 17.1% of the graft weight.
- Inter-Reader Agreement: Nearly perfect agreement was observed between two readers for graft volume measurement (CCC = 0.998), highlighting the consistency achievable with the DLA-assisted approach, even with varying levels of experience.
Limitations and Future Work
The authors acknowledge several limitations:
- Retrospective Design: The study's retrospective nature may introduce selection bias and potential biases from missing data.
- Single Institution: The development and validation groups were from the same institution, suggesting that external validation in a different population would be beneficial for broader applicability.
- Right-Lobe Grafts Only: The study focused exclusively on right-lobe grafts, which are preferred for LDLT. The findings may not directly apply to other types of liver grafts.
- Hepatic Steatosis (HS): While HS did not show a significant confounding effect in the development group, the study population had predominantly mild or no HS. Further investigation is needed for donors with moderate-to-severe HS.
Future work could involve external validation of the developed formula and DLA in diverse populations and exploring its applicability to other types of liver grafts. Investigating the impact of HS more thoroughly on graft weight estimation would also be valuable.
Summary updated