A Deep Learning Harmonization of Multi-Vendor MRI for Robust Intervertebral Disc Segmentation

Abstract

Magnetic resonance imaging (MRI) provides enhanced soft tissue contrast and high spatial resolution. However, the relationship between intensity values among soft tissues in MRI is inconsistent, even when obtained under the same conditions (e.g., vendors and acquisition protocols). This inconsistency hinders accurate medical image segmentation and disease classification. Therefore, we propose a framework to harmonize multi-vendor MRI using a novel radiomics approach for robust segmentation. The proposed model comprises a cycle-consistent adversarial network (CycleGAN)-based network and a segmentation network. The CycleGAN-based network harmonizes MRI with the support of a radiomics-based method (radiomic feature (RF) loss function newly designed for this study). The segmentation network encourages the CycleGAN-based network to enhance intervertebral disc (IVD) segmentation features using dice loss functions during harmonization. Furthermore, publicly available datasets and diverse MRI scans provided by a collaborating hospital were used to make our model more robust to MRI variability. The proposed model was evaluated for segmentation using the Dice coefficient, intersection-over-union (IoU), F1 score, precision, and recall. It outperformed other segmentation methods (Dice = 0.920, IoU = 0.853, F1 score = 0.920, precision = 0.940, and recall = 0.902), even on diverse test datasets with disease information. The harmonization performance was assessed using the relative error of the RF values between the target (standard) and harmonized data. It achieved the four best scores ( $\mathrm {\approx 0}$ ) among the five features in a relative error of RF compared to other harmonization methods (e.g., conventional histogram-based method and deep learning model).

Keywords

• Harmonization
• magnetic resonance imaging
• radiomics
• segmentation