Evaluation of conventional and deep learning based image harmonization methods in radiomics studies

Area of Science:

• Radiomics
• Medical Imaging
• Artificial Intelligence

Background:

• Radiomic features extracted from medical images can predict patient outcomes.
• Variations in imaging protocols across institutions can introduce biases, affecting the reliability of radiomic models.
• Image harmonization techniques aim to standardize image properties, enhancing model generalizability.

Purpose of the Study:

• To evaluate the impact of image harmonization on radiomic feature extraction and outcome prediction models.
• To compare the effectiveness of histogram matching (H_HM) and generative adversarial network (GAN)-based (H_GAN) harmonization methods.
• To assess the influence of harmonization on the selection and contribution of different radiomic feature types (shape, histogram, texture).

Main Methods:

• T1-weighted MRI scans from 234 patients in the Brain Tumor Image Segmentation Benchmark (BRATS) dataset were harmonized using H_HM and H_GAN.
• 88 radiomic features were extracted from original (H_NONE), H_HM, and H_GAN images.
• Statistical analysis (Wilcoxon paired test) identified features significantly altered by harmonization.
• Radiomic prediction models were built using Least Absolute Shrinkage and Selection Operator (LASSO) for feature selection and Kaplan-Meier analysis for survival prediction.

Main Results:

• Over 50% of radiomic features (49/88 with H_HM, 55/88 with H_GAN) were significantly modified by harmonization (adjusted p < 0.05).
• Harmonized datasets showed an increased contribution of histogram and texture features to LASSO-selected models compared to shape features.
• Both H_HM and H_GAN enabled the creation of survival prediction models with significantly different patient survival groups (p < 0.05).
• The H_GAN method allowed building a validated model using only harmonization-impacted features, showing distinct median survivals (189 vs. 437 days, p = 0.006).

Conclusions:

• Data harmonization in multi-institutional cohorts can recover the predictive value of radiomic features affected by inter-center imaging property variations.
• GAN-based harmonization (H_GAN) outperformed histogram matching (H_HM) in building survival prediction models, demonstrating its potential for advanced radiomic analysis.
• Harmonization techniques are crucial for enhancing the robustness and clinical applicability of radiomic models, particularly in multi-center studies.