Abstract
Background: This study aims to explore the relationship between early allograft dysfunction (EAD) and the visual appearance of the liver post-reperfusion, and to develop image-based models for predicting EAD and short-term mortality.
Methods: A total of 351 liver transplant recipients were enrolled and divided into training (n=229) and testing (n=122) sets. RGB images of donor livers were captured post-reperfusion. Support vector machines (SVM) and convolution neural network (CNN) models based on texture analysis were constructed to predict EAD. The best-performing model was further refined with additional clinical information. An "image score" was assigned based on CNN prediction probabilities to evaluate correlations with clinical outcomes.
Results: The CNN model achieved an AUC of 0.709 in the testing set, outperforming the SVM model (AUC 0.661). A hybrid model incorporating clinical data reached an AUC of 0.727. Higher image scores were significantly associated with increased postoperative infusions, more complications, and longer ICU and hospital stays.
Conclusion: Post-reperfusion donor liver appearance is associated with EAD occurrence. Texture analysis of RGB images provides a feasible, non-invasive method for predicting EAD and patient outcomes.
Introduction
Early Allograft Dysfunction (EAD) is a critical "gray zone" in liver transplantation: the new liver functions enough to keep the patient alive but shows signs of significant damage. EAD is a major predictor of recipient morbidity and mortality, often leading to longer hospital stays and increased susceptibility to sepsis.
Currently, surgeons often assess the health of a liver graft through subjective visual inspection—looking at its color and texture after blood flow is restored (post-reperfusion). While experienced surgeons have a "gut feeling" about graft quality, this isn't quantifiable or standardized. Existing objective methods often require specialized equipment or rely solely on donor/recipient clinical data, which may not capture the real-time physiological state of the liver after surgical trauma.
Methodology
Researchers explored two primary computational paths to turn visual "gut feelings" into objective data:
- Classical Machine Learning (SVM): This approach used traditional Texture Analysis (TA), extracting "hand-crafted" features such as Local Binary Patterns (LBP), Gray Level Co-occurrence Matrices (GLCM), and basic intensity metrics. These features were then processed by a Support Vector Machine (SVM).
- Deep Learning (CNN): A Convolutional Neural Network (CNN) inspired by the Google Inception-Net architecture was developed. Unlike the SVM, the CNN automatically learns relevant visual features directly from RGB images.
The best-performing model (CNN) was then used to create a hybrid version incorporating three key clinical variables: donor age, surgery time, and cold ischemia time. Finally, the CNN’s prediction probability was translated into a simplified "Image Score" (1 to 3) to predict broader clinical outcomes.
Data Overview
The study utilized a dataset of 351 liver transplant recipients from the First Affiliated Hospital, Sun Yat-sen University (2017–2019). The training set included 229 patients, while the testing set included 122. Input data consisted of RGB images captured using standard smartphones in the operating room before abdominal closure, alongside electronic medical records.
Key Insights
- Smartphone as a Diagnostic Tool: By using standard smartphone cameras, the researchers demonstrate a path toward high-tech diagnostics that require zero additional equipment in the operating room.
- Texture as a Proxy for Pathology: The study confirms that the "look" of a liver post-reperfusion—its micro-texture and color distribution—contains quantifiable signals of ischemia-reperfusion injury detectable by AI.
- Beyond Binary Prediction: The "Image Score" correlated significantly with the intensity of recovery. Patients with higher (worse) scores required more postoperative infusions and stayed longer in the ICU, suggesting the model captures a spectrum of graft health.
Results and Clinical Implications
- CNN vs. SVM: The CNN outperformed the classical SVM approach, achieving an Area Under the Curve (AUC) of 0.709 compared to 0.661 in the testing set.
- The Hybrid Advantage: Integrating clinical information with CNN features improved predictive power to an AUC of 0.727.
- Clinical Correlation: A higher Image Score was statistically significant in predicting increased postoperative complications (e.g., hydrothorax, bleeding), longer ICU and hospital stays, and higher volumes of required blood product transfusions.
Limitations and Future Directions
Several hurdles remain before routine clinical application. The study was conducted at a single center, necessitating multi-center validation. Image heterogeneity due to varying lighting and smartphone models introduces noise; standardizing acquisition could improve accuracy. Furthermore, while 351 patients is a significant cohort, deep learning models benefit from larger datasets to improve sensitivity, which was relatively low (54%) in the final model. Finally, the "black box" nature of CNNs makes it challenging to pinpoint exactly which visual features the model prioritizes.
Summary updated