Multi-parametric MRI-based machine learning model for prediction of pathological grade of renal injury in a rat kidney cold ischemia-reperfusion injury model

Lihua Chen, Yan Ren, Yizhong Yuan, Jipan Xu, Baole Wen, Shuangshuang Xie, Jinxia Zhu, Wenshuo Li, Xiaoli Gong, Wen ShenBMC Medical Imaging 2024 · DOI

TL;DR

This study introduces a novel machine learning model that uses multi-parametric MRI to non-invasively predict the severity of kidney injury in a rat model of cold ischemia-reperfusion injury (CIRI). The model achieves high accuracy, offering a promising alternative to invasive biopsy for early and quantitative assessment of renal damage.

What Problem Does This Paper Address?

Kidney cold ischemia-reperfusion injury (CIRI) is a critical complication in kidney transplantation, often leading to delayed graft function and reduced long-term graft survival. Currently, accurately assessing the degree of CIRI is challenging. While biomarkers exist, they are not always used clinically. The gold standard, kidney biopsy, is invasive and carries risks. Existing non-invasive methods like serum creatinine and blood urea nitrogen are not sensitive enough for early detection. This paper aims to bridge this gap by developing a non-invasive, quantitative method for predicting CIRI severity using advanced MRI techniques and machine learning.

Methods

The researchers employed a multi-parametric MRI approach, acquiring data from various sequences including Diffusion-Weighted Imaging (DWI), Intravoxel Incoherent Motion (IVIM), Diffusion Kurtosis Imaging (DKI), Blood Oxygen Level-Dependent (BOLD), T1 mapping, and Arterial Spin Labeling (ASL). From these scans, 25 distinct parameters related to renal cortex and medulla were extracted.

To handle the complexity and potential redundancy of these parameters, a two-stage machine learning process was used:

  1. Feature Selection: The Least Absolute Shrinkage and Selection Operator (LASSO) regression, combined with five-fold cross-validation, was utilized to identify the most relevant features for predicting injury grade. This method helps in selecting a parsimonious set of predictors while preventing overfitting.
  2. Model Construction: After feature selection, several machine learning classifiers were trained and evaluated, including Neural Network (NN), Support Vector Machine (SVM), Random Forest (RF), multinomial logistic regression, Bayes, Decision Tree, and K Nearest Neighbors (KNN). The performance of these models was assessed using metrics like accuracy, Area Under the Curve (AUC), and F1-score.

Histopathological analysis, using Hematoxylin and Eosin (HE) staining and a semi-quantitative scoring system, served as the ground truth for classifying the severity of renal injury into mild, moderate, and severe groups via K-means clustering.

Datasets

The study utilized data from 80 male Sprague Dawley rats, divided into four groups: sham operation, and 1, 2, and 4 hours of cold ischemia. MRI scans were performed on these rats at specific time points post-operation. The histopathological scores from these rats formed the ground truth dataset for training and evaluating the machine learning models. No new public benchmark datasets were introduced.

Key Insights and Ideas

  • Synergy of Multi-parametric MRI: The study highlights that combining information from multiple MRI sequences provides a more comprehensive picture of renal injury than any single technique. Different MRI parameters capture distinct aspects of tissue damage, such as microvascular changes, water diffusion, and oxygen metabolism.
  • IVIM and BOLD Shine: Among the various MRI techniques, IVIM and BOLD showed the highest efficacy in grading renal injury. This suggests that parameters related to microcirculation (IVIM) and tissue oxygenation (BOLD) are particularly sensitive to CIRI.
  • Machine Learning as a Powerful Integrator: Machine learning, specifically logistic regression, RF, and SVM, proved effective in integrating these diverse MRI biomarkers into a single, predictive classifier. This demonstrates the potential of AI in translating complex imaging data into clinically relevant assessments.
  • Non-invasive Prediction is Feasible: The successful development of a predictive model based on MRI parameters offers a viable non-invasive alternative to kidney biopsy for assessing CIRI severity, which could lead to earlier interventions and improved patient outcomes.

Key Results

  • The study identified 8 key features that were highly correlated with pathological grading: medullary Dp, cortical Dp, cortical T2, cortical Fp, medullary T2, ΔT1, cortical RBF, and medullary T1.
  • Multinomial logistic regression, Random Forest, and SVM models demonstrated significantly higher accuracy and AUC compared to other classifiers.
  • The multinomial logistic regression model achieved the highest AUC (0.95) for pathological grading.
  • When comparing individual MRI techniques for grading, IVIM yielded the highest AUC (0.93), followed closely by BOLD (0.90).
  • Internal validation showed the predictive model's robustness, with accurate probability predictions for different injury grades.

Limitations and Future Work

The authors acknowledge several limitations:

  • Single Injury Model: The study focused on a single type of CIRI model in rats. Future work should explore models with different injury mechanisms and a larger number of animals to improve model generalizability.
  • Single-Center Study: The research was conducted at a single center. External validation with larger, multi-center datasets and diverse patient populations is crucial for broader clinical application.
  • Parameter Maps Only: The current model relies on parameter maps derived from MRI. Investigating the utility of routine T2WI sequences, texture features, or radiomic features could potentially further enhance prediction accuracy.
  • Data Imbalance: While cross-validation was used, the authors note that data imbalance is a common challenge in classification research, and further exploration of techniques to address this is warranted.

Future work should focus on developing more stable classification models, incorporating external validation, and exploring the integration of additional imaging features to improve the prediction of renal damage severity in patients.

Summary updated

BibTeX
@article{chen2024multi,
  author = {Lihua Chen and Yan Ren and Yizhong Yuan and Jipan Xu and Baole Wen and Shuangshuang Xie and Jinxia Zhu and Wenshuo Li and Xiaoli Gong and Wen Shen},
  title = {{Multi-parametric MRI-based machine learning model for prediction of pathological grade of renal injury in a rat kidney cold ischemia-reperfusion injury model}},
  journal = {BMC Medical Imaging},
  year = {2024},
  doi = {10.1186/s12880-024-01320-6},
}
Figure