Radiomics-based kidney lesion classification: Mitigating batch effect with nested combat harmonization.

Area of Science:

• Medical Imaging Analysis
• Computational Pathology
• Machine Learning in Oncology

Background:

• Increasing CT scans lead to more incidental renal masses, requiring differentiation between benign and malignant types.
• Radiomics shows promise for improved renal mass diagnosis but is hindered by imaging parameter variability, like slice thickness.
• Effective harmonization techniques are essential to standardize radiomics data and improve diagnostic reliability.

Purpose of the Study:

• To perform a comprehensive radiomics analysis to evaluate the impact of slice thickness on distinguishing kidney cysts from tumors.
• To leverage machine learning techniques for improved accuracy in renal mass classification.
• To contribute to more precise patient management strategies for renal masses.

Main Methods:

• Utilized the KITS23 dataset of 599 contrast-enhanced CT scans, split into training (60%) and testing (40%) cohorts.
• Extracted radiomic features using PyRadiomics, applying six feature selection methods and ten machine learning classifiers.
• Implemented the Nested Combat harmonization technique to address inter-institutional imaging protocol variations.

Main Results:

• Harmonization using Nested Combat led to improved Area Under the Curve (AUC) values across various methods and classifiers.
• The highest achieved AUC reached 0.95, demonstrating significant model performance enhancement.
• Mean AUC improvements ranged from 0.7% to 7.7%, with results comparable to or exceeding existing literature benchmarks.

Conclusions:

• Radiomics-based machine learning models hold significant potential for enhancing diagnostic accuracy in renal oncology.
• Harmonization techniques, specifically Nested Combat, are critical for developing reliable and generalizable predictive models.
• These advancements can lead to improved patient management strategies in clinical practice.