Multiparametric MRI

From Simultaneous Rapid Acquisition Methods and Analysis Techniques Using Scoring, Machine Learning, Radiomics, and Deep Learning to the Generation of Novel Metrics

Abstract

With the recent advancements in rapid imaging methods, higher numbers of contrasts and quantitative parameters can be acquired in less and less time. Some acquisition models simultaneously obtain multiparametric images and quantitative maps to reduce scan times and avoid potential issues associated with the registration of different images. Multiparametric magnetic resonance imaging (MRI) has the potential to provide complementary information on a target lesion and thus overcome the limitations of individual techniques. In this review, we introduce methods to acquire multiparametric MRI data in a clinically feasible scan time with a particular focus on simultaneous acquisition techniques, and we discuss how multiparametric MRI data can be analyzed as a whole rather than each parameter separately. Such data analysis approaches include clinical scoring systems, machine learning, radiomics, and deep learning. Other techniques combine multiple images to create new quantitative maps associated with meaningful aspects of human biology. They include the magnetic resonance g-ratio, the inner to the outer diameter of a nerve fiber, and the aerobic glycolytic index, which captures the metabolic status of tumor tissues.

With the recent advancements of magnetic resonance imaging (MRI), acquiring multiple contrast-weighted images has become clinical routine. For specific purposes such as cancer detection and investigation of age-related changes, a series of magnetic resonance (MR) images, collectively called multiparametric MRI, has been used to investigate normal and pathological tissues.^1–4 Some acquisition models, including synthetic MRI^5–8 and MR fingerprinting,^9–11 have been developed to simultaneously acquire multiparametric images to reduce scan time and avoid potential issues associated with the registration of different images. These techniques and a further improvement in acceleration techniques are expected to boost the number of images acquired in a single clinical scan session, likely increasing the necessity of automatic analysis methods for the acquired data. Multiparametric MRI has the potential to provide complementary information about a target lesion and thus overcome the limitations of individual techniques. Various techniques for the analysis of multiparametric MRI data are already available. Scoring systems are widely used in clinical settings and ensure high interpretability and usability.¹² Machine learning models, including radiomic analysis and deep learning, have been gaining popularity in research settings in recent years. Still, their biases and relatively low generalizability hinder their introduction into clinical practice.^13,14 Another unique approach to analyzing multiparametric MRI data is to create new quantitative maps associated with some meaningful in vivo biological aspects.

This review article aims to summarize and discuss the acquisition methods of multiparametric MRI, with a special focus on simultaneous acquisition techniques, and how multiparametric MRI data can be analyzed as a whole rather than each parameter separately.

ACQUISITION OF MULTIPARAMETRIC MR IMAGES

Developing efficient acquisition techniques such as turbo spin-echo and gradient echo, as well as acceleration techniques such as parallel imaging¹⁵ and compressed sensing,¹⁶ have enabled the acquisition of multiple MR images in a clinically acceptable scan time.¹⁷ Recently, quantitative imaging has been developing fast, and various methods, which are roughly categorized into indirect and direct measurement methods,¹⁸ can be used to acquire T1 and T2 relaxation times and proton density maps simultaneously (Fig. 1). Indirect methods, such as driven-equilibrium single-pulse observation of T1 (DESPOT1) and T2 (DESPOT2),^19,20 2-dimensional quantification of relaxation times and proton density by multiecho acquisition of a saturation-recovery using Turbo spin-Echo Readout (2D-QRAPMASTER),^5,6 and 3D quantification using an interleaved Look-Locker acquisition sequence with a T2 preparation pulse (3D-QALAS),^21,22 simultaneously acquire multiple images with different contrast weightings and calculate quantitative values via voxel-by-voxel fitting of the equations based on the acquisition parameters of the pulse sequence. Direct methods include MR fingerprinting⁹ and MR multitasking,²³ which acquire data fragments that are used to generate quantitative maps without producing intermediate contrast-weighted images. Although implementation dependent, it is now possible to obtain 3D whole-brain relaxation maps using indirect or direct methods in less than 6 minutes. In addition to T1 and T2 relaxation times and proton density, some methods also enable simultaneous measurements of diffusion,^23,24 susceptibility,²⁵ perfusion,²⁶ magnetization transfer,^27,28 and myelin.^29,30

FIGURE 1: Overview of simultaneous multiparametric MRI acquisition. Acquisition approaches can be categorized into (A) indirect and (B) direct mapping methods based on whether or not intermediate fully reconstructed contrast-weighted images are produced. 3D-QALAS indicates 3D-quantification using an interleaved Look-Locker acquisition sequence with a T2 preparation pulse; MRI, magnetic resonance imaging.

Advanced reconstruction methods can be applied to multiparametric methods to improve the image quality or reduce the scan time. Because simultaneous multiparametric imaging methods are compatible with most reconstruction algorithms that utilize data redundancy and consistency, various techniques have been applied to simultaneous multiparametric imaging.³¹ These acquired images can be passed to the data fitting step to efficiently generate quantitative maps. Image reconstruction for multiparametric mapping is an active field of research. Recent advances, such as low-rank subspace modeling and model-based reconstruction,³¹ have accelerated multiparametric MRI,^32–36 potentially improving workflow and patient comfort. Low-rank modeling takes advantage of efficient representation of medical images to reduce degrees of freedom and allows the recovery of missing information from sparsely sampled data. Low-rank image reconstruction is especially powerful in recovering high-dimensional images, making it suitable for multicontrast images (thus multiparametric mapping) and dynamic images. Model-based reconstruction is an approach that directly maps from the k-space with an iterative process. This ensures data consistency of the final map with the acquired k-space data. For detailed descriptions, please refer to Akcakaya et al.³¹

Although tissue relaxivities are unspecific to the underlying pathology, incorporating tissue relaxation times in other MRI sequences has been shown to add value for diagnosing diseases such as prostate cancer^37,38 and Alzheimer disease.^39,40

A set of multiple parameters could be used to “synthesize” a set of contrast-weighted images for diagnostic purposes. Because clinicians are not accustomed to visual analysis of quantitative maps, making a specific diagnosis solely on quantitative maps is difficult. With synthetic MRI applied to simultaneous relaxometry,^5,6,41–44 contrast-weighted images, such as T1-weighted, T2-weighted, and fluid-attenuated inversion recovery (FLAIR) images, can be created based on T1, T2, and proton density maps without misregistration issues. The postprocessing parameters (ie, repetition time, echo time, and inversion time) could be easily extended to ranges (eg, minimal echo time) that are technically challenging by conventional acquisition, thereby facilitating the extraction of high-quality diagnostic information. Synthetic MRI enables multiparametric MRI to be incorporated into clinical practice without additional scan time.

A limitation of incorporating a novel technique into multiparametric imaging is the paucity of established frameworks for acquisition, evidence for standardization and variability in the image, and clear thresholds for abnormality, leading to increased uncertainty in the outcome for combinations of individual modalities.^45,46 Another drawback of multiparametric mapping techniques is the vulnerability to subject motion during the scan.⁴⁷ Because data for all maps are acquired simultaneously, subject motion at any time during the acquisition will degrade all resulting maps. Several techniques have been proposed to make the acquisition robust to patient motion.^48–52

Although multiparametric imaging techniques have the potential to allow a more comprehensive characterization of pathological processes and therapeutic responses, the stability of the quantitative maps is a prerequisite for the acquired maps to be used in multi-institutional studies. The repeatability and reproducibility of the maps acquired with multiparametric imaging techniques are currently under evaluation, usually in multisite and multiplatform settings.^53–58 The standardization of the acquisition protocols is an active field of research. The quantitative MRI techniques referenced in this section have the potential to offer more reproducible results compared with conventional contrast-weighted images owing to their quantitative nature. For a discussion related to the variability and standardization of multiparametric MRI, the reader can refer to a review article by Hagiwara et al.⁴⁶

ANALYZING MULTIPARAMETRIC MR IMAGES

Representative techniques to analyze multiparametric data in the MRI field are summarized in Figure 2. In clinical settings, clinicians visually assess multiple MR images and make a diagnosis. Scoring systems using multiparametric MRI data, such as Reporting and Data Systems (RADS), have recently come into clinical practice for the clinical management of cancers. Meanwhile, mathematical analysis, such as machine learning and deep learning, of multiparametric MR images has been mainly performed in research settings. The MRI techniques referenced in this section are mainly conventional contrast-weighted images and quantitative MRI approaches, such as diffusion-weighted imaging (DWI) and perfusion MRI. Simultaneous multiparametric imaging is still preliminary and the literature analyzing such data is scarce. These examples are also presented in this section.

FIGURE 2: Representative methods for analyzing multiparametric MRI data along with their advantages and disadvantages.

Scoring Systems

Reporting and Data Systems are guidelines for assessing region- and disease-oriented imaging studies. Since the prototype Breast Imaging RADS (BI-RADS) was first published in 1993, many standardized reporting systems have been developed. As of late 2022, there are 5 American College of Radiology–endorsed RADS using MRI features for scoring as follows: BI-RADS, Liver Imaging RADS (LI-RADS), Musculoskeletal RADS (MSK-RADS), Neck Imaging RADS (NI-RADS), and Prostate Imaging RADS (PI-RADS).¹² These RADS are modality- and sequence-specific and designed to minimize terminology heterogeneity and discrepancies in malignant risk assessment among interpreters. Given their systematic nature, the introduction of RADS can also be expected for applications in consistent data collections⁵⁹ and future deep-learning model designs.^1,60,61 Most RADS acknowledge the usefulness of multiparametric MRI. For example, in a multicenter study, Clauser et al⁶² found that applying an apparent diffusion coefficient (ADC) cutoff of ≥1.5 × 10⁻³ mm²/s in addition to contrast-enhanced MRI allowed the downgrading of lesions classified as BI-RADS 4 (2%–95% likelihood of malignancy), which potentially avoided 32.6% of unnecessary biopsies while maintaining a high sensitivity of 96.6%. In a meta-analysis including 16 studies between 2016 and 2020,⁶³ the pooled sensitivity for diagnosing hepatocellular carcinoma using LI-RADS was the highest with arterial phase hyperenhancement, and the pooled specificity was the highest for enhancing capsule detected in the portal venous phase or delayed phase; thus, dynamic phases should be evaluated in combination.⁶⁴ Elsholtz et al⁶⁵ found that for the primary site, interreader agreement on NI-RADS scores of contrast-enhanced MR images was significantly lower than that on DWI, which is currently not part of the NI-RADS criteria; thus, they concluded that DWI could augment the NI-RADS reliability.

Low-field MRI benefits from reduced susceptibility effects and is a promising alternative for lung imaging to avoid radiation exposure due to computed tomography.^66,67 Lévy et al suggested a scoring system using perfusion and ventilation function metrics derived from a phase-resolved functional lung technique with a free-breathing 0.55-T MRI.⁶⁸ They applied the scoring system to patients with persistent symptoms after COVID-19 infection, and the combined score differed between patients with and without persistent symptoms. The combined score also showed a consistent increase from healthy volunteers via patients without persistent symptoms to those with persistent symptoms.

We should be mindful that scoring systems come with unwanted interobserver variability.^69,70 There are ongoing efforts to automate the process of detecting lesions and differentiating benign from malignant lesions, showing a comparable or superior performance compared with expert radiologists.^71–73

Machine Learning and Radiomics

Artificial intelligence is an umbrella term, and machine learning is the most popular and most applied subcategory in radiology (Fig. 3).⁷⁴ Artificial intelligence is the simulation of human intelligence by computer algorithms, whereas machine learning is a way of learning from data or experience to find generalizable predictive patterns, leading to automation of analytical model building. Recent studies have exploited machine learning in evaluating multiparametric MRI data partly because machine learning is particularly effective when applied to broad data where the number of input variables exceeds the number of subjects.⁷⁵ Radiomics has become one of the most popular analytical methods of multiparametric MRI, which is based on the extraction of potentially innumerable numbers of quantitative imaging metrics, or “radiomic features,” which are difficult to be captured by the human eye.⁷⁶ Radiomic features are collectively used for the prediction of diagnosis,^77–80 classification,³⁷ prognostication,⁸¹ assessment of treatment effects,⁸² and gene expression profiling, among others. Focal lesions (eg, tumors) or other areas of concern are segmented to be used as regions of interest for extracting radiomic features.

FIGURE 3: Relationships between artificial intelligence and its subcategories.

Numerous studies have found the usefulness of multiparametric MRI using multivariate logistic regression models to establish clinical and radiological parameters that predict malignancies and clinically relevant risk factors. For example, in differentiating transition zone prostate cancer and benign prostate hyperplasia, Iyama et al⁸³ found that standardized T2-weighted imaging signal and mean ADC were independent factors for the differential diagnosis. The combined model offered an area under the receiver operating characteristics curve (AUC) of 0.98, comparable with board-certified radiologists' diagnoses and superior to PI-RADS. Chawla et al⁸⁴ combined median ADC and dynamic contrast-enhanced (DCE) MRI parameters and distinguished chemoradiation responders from nonresponders among patients with head and neck squamous cell carcinoma, with an AUC of 0.85. For predicting microvascular invasion of hepatocellular carcinoma, which is an adverse prognostic factor, Liao et al⁸⁵ and Wang et al⁸⁶ reported high predictive values of multiparametric MRI using multivariate logistic regression analyses. The AUC was 0.84 with the model using tumor size (≥3 cm), single tumor involving more than 2 segments, mean ADC, and minimum ADC in the study by Liao et al,⁸⁵ whereas the AUC was 0.784 by combining mean kurtosis and irregular circumferential enhancement in the study by Wang et al.⁸⁶ Another machine learning study with linear discrimination analysis utilized T1, T2, and proton density values acquired with QRAPMASTER extracted from gray matter, white matter, and lesion masks.⁸⁷ They used mean values of these masks to differentiate between multiple sclerosis, a demyelinating disorder of the central nervous system, and hereditary diffuse leukoencephalopathy with spheroids, a rare demyelinating disorder that can sometimes be misdiagnosed as multiple sclerosis. They achieved 100% accuracy, although their small cohort may not have fully represented the variability of the features associated with these diseases

Radiomics using MRI has been typically performed using conventional T1-weighted, T2-weighted, and FLAIR images. For example, Ismail et al⁸⁸ used routinely available contrast-enhanced T1-weighted and FLAIR images to extract 60 shape features of glioblastoma. They applied a feature selection algorithm to reduce them to 5 and used a support vector machine to differentiate true progression from pseudoprogression. Contrast-enhancing tumors with true progression were round and compact compared with those with pseudoprogression; FLAIR hyperintense tumors with true progression were rounder. Quantitative imaging, such as ADC,⁸⁹ dynamic susceptibility contrast (DSC) perfusion,⁸⁹ chemical exchange saturation transfer (CEST),⁹⁰ magnetization transfer imaging,⁹¹ and even positron emission tomography (PET),⁹² can also be added to routine MRI for radiomic analysis. Kim et al⁸⁹ used contrast-enhanced T1-weighted and FLAIR images, cerebral blood volume (CBV), and ADC to develop a multiparametric radiomic model that differentiates between true progression and pseudoprogression in patients with glioblastoma. As a result, the multiparametric radiomics model (AUC, 0.90) showed significantly better performance than any one of the parameters mean or maximum CBV or mean or minimum ADC (AUC, 0.57–0.79) and had a higher AUC than radiomic models using conventional MRI (AUC, 0.76), CBV (AUC, 0.80), or ADC (AUC, 0.78) alone. The multiparametric radiomic model was validated in an external dataset from another hospital with different acquisition parameters from the development dataset, with a reduced AUC of 0.85. This is partly associated with variations in how DSC was performed in different hospitals, such as low and high flip angles.⁹³ This is one reason multiparametric techniques can be challenging to translate into clinical practice. Furthermore, when using an imaging technique that is not routinely performed, it would be challenging to gather internal and external datasets with sufficient subject numbers for model development and validation.

The reproducibility of radiomic studies still needs to be improved, representing obstacles to clinical implementation.^46,94 Variations in any step of the radiomics workflow can influence feature values and the final output, including acquisition parameters, reconstruction algorithms, image processing steps, region of interest placements, and feature selection methods.^95–98 There is no clear consensus on the optimal number of MRI sequences to be included in a radiomic model. A phantom study can be performed to investigate the repeatability and reproducibility of the radiomic features of each MRI sequence.^46,99 This information can be used to select MRI sequences into a radiomic model before launching a complete radiomic study.

Radiomics has been primarily applied for tumor evaluation. However, recent studies have developed radiomic models for cerebral infarction,¹⁰⁰ epilepsy,^101–103 and multiple sclerosis.^91,104–106 However, this field seems to be in its preliminary stage. Most of these studies used only a single contrast image^104,105 or a quantitative map, such as magnetization transfer imaging⁹¹ and quantitative susceptibility mapping,¹⁰⁶ for example, to differentiate between multiple sclerosis and ischemic disease¹⁰⁴ or neuromyelitis optica spectrum disorder¹⁰⁶ and predict whether a focal multiple sclerosis lesion progresses or not.¹⁰⁵ Quan et al¹⁰⁰ performed radiomic analysis on FLAIR images and ADC maps to predict unfavorable outcomes (modified Rankin Scale score >2) of acute ischemic stroke and showed that the combined model of FLAIR and ADC was superior to the model using either of them and the models using clinical information, conventional MRI, or both.

Recently, many studies using radiomics have been published, and a radiomics quality score (RQS) was developed to assess the quality of radiomic studies.^107,108 Park et al. evaluated the quality of 77 radiomic studies in oncology using RQS and the Transparent Reporting of a Multivariable Prediction model for Individual Prognosis or Diagnosis (TRIPOD) statement, guidelines for reporting of prediction models for prognosis or diagnosis. These authors determined that the mean RQS and TRIPOD adherence rate were 26.1% and 57.8%, respectively, of the maximums and concluded that overall scientific quality and reporting of radiomics studies are insufficient.¹⁰⁹ The RQS was particularly low in demonstrations of clinical utility, test-retest analyses, prospective studies, and open science.¹⁰⁹ In the RQS, validation of the radiomic model using multiple external datasets is given one of the highest scores. It is crucial to ensure reproducibility and generalizability for a machine learning model to be used in clinical settings. However, a recent systematic review demonstrated that only 29.4% of the radiomic studies in the neuro-oncology field had validated their radiomic model using external validation datasets.¹¹⁰ Many radiomic studies seem to lack substantial evidence. Nonetheless, the broad potential of radiomics has been shown.

Habitat Imaging

Habitat imaging is another machine learning technique that handles multiparametric MRI data simultaneously. It divides imaging data into distinctly different segments and can provide unique insights into associations between multiparametric MRI and biological subpopulations, or habitats, of a tumor.¹¹¹ This is typically performed via unsupervised learning, such as K-means clustering and Gaussian mixture models, for the image voxels to form distinct clusters. One of the limitations of other machine learning techniques is that they assume only 1 pathology, and the presence of more than 1 pathology (eg, an abscess associated with a tumor) can confound algorithms because these cases are scarce and difficult to label correctly.⁴⁵ Habitat imaging does not assume 1 pathology in a lesion and may enable it to overcome this limitation. Furthermore, statistical models and other machine learning techniques often ignore spatial information. Habitat imaging can combine multiparametric MRI data into a single habitat map. However, it requires spatial registration of all images and must consider differences in resolution and deformations that can occur between different scans.^111,112

Preclinical studies have advantages in coregistering habitats identified by MRI with histology. For example, T2 and T2* maps, DWI, and DCE-MRI of a breast cancer mouse model were combined to identify 4 different habitats that were initially classified on histology, namely, viable-normoxic, viable-hypoxic, nonviable-hypoxic, and nonviable-normoxic (Fig. 4).^111,113 A 3D-printed tumor mold was used to aid the coregistration of the histological slices with MRI, resulting in a good visual agreement between the habitat map derived from multiparametric MRI and the corresponding histological slice.

FIGURE 4: Multiparametric MRI data used to identify tumor subregions (“habitats”) in a 4 T1 breast cancer allograft in a mouse model. (A) Parameters derived from T2 map, T2* map, DWI-MRI, and DCE-MRI were clustered to create a habitat map, demonstrating 4 intratumor subpopulations with different patterns of cellularity. Blue indicates necrotic; yellow, nonviable hypoxic; pink, viable hypoxic; green, viable nonhypoxic. (B) A coregistered habitat map derived from histology demonstrating good spatial correspondence with the habitat map derived from MRI. Histological slices were obtained by cutting the tumors using a 3D printed mold, which ensured the coregistration with MRI slices. Hematoxylin and eosin and immunohistochemistry images were automatically segmented and superimposed to create habitat maps derived from histology, showing necrosis (blue), viable normoxic (green), pimonidazole in a viable region (pink), and pimonidazole in a nonviable region (yellow). Reproduced with permission from Napel et al.¹¹¹

Clinically, habitats identified by MRI have been used for prediction of genetic mutation,^114–116 classification,¹¹⁷ prognostication,^118–122 and assessment of treatment effect.^123–127 For example, Hagiwara et al¹¹⁴ combined contrast-enhanced T1-weighted image, FLAIR image, amine CEST, and ADC of human gliomas and used unsupervised clustering to label each voxel to acquire habitat maps. They categorized tumor voxels that are typical of isocitrate dehydrogenase (IDH) mutant glioma, IDH wild-type glioma, or neither and used this information to differentiate between IDH mutant and wild-type gliomas. Furthermore, they correlated these labels with immunohistochemistry based on MRI-guided biopsy and demonstrated that tumor habitats identified by multiparametric MRI captured the difference in metabolic profiles of IDH mutant and wild-type gliomas (Fig. 5).

FIGURE 5: MR images and corresponding hematoxylin and eosin (H&E) and immunohistochemistry staining for MRI-guided biopsy targets (circles). (a) IDH mutant glioma for which an area with labels categorized as M, indicating the IDH mutant feature, was biopsied. Expressions of hypoxia-inducible factor 1-α (HIF1a), glucose transporter 3 (GLUT3), and hexokinase 2 (HK2) are low in the slides from a 5-mm radius sample taken from the MRI-guided biopsy target. (b) IDH wild-type glioma for which an area with labels categorized as W, indicating IDH wild-type feature, was biopsied. Expressions of HIF1a, GLUT3, and HK2 are high. HIF1a, GLUT3, and HK2 are known to be elevated in IDH wild-type gliomas. Reproduced with permission from Hagiwara et al.¹¹⁴

Deep Learning

Deep learning is a subset of machine learning that is also a subset of artificial intelligence (Fig. 3). Deep learning refers to a data-driven learning approach that computes using multilayer neural networks.⁷⁴ Each layer can be viewed as a learnable function that performs a specific calculation and produces an output that serves as the input for the next layer. Convolutional neural network (CNN) is a type of deep learning known to be useful for image tasks, with its architecture resembling that of the human visual cortex.^128,129 In recent years, deep learning, particularly CNN, has gained substantial popularity in the MRI field for image reconstruction, image quality improvement, image transfer, disease detection, tissue segmentation, and classification.^130–142

Multiparametric MRI has been used for analyses using deep learning. Multiparametric MRI is potentially useful for lesion segmentation to fully capture the extent of the disease and has been used to segment cancers and multiple sclerosis plaques.^{133,139,141,143–149} Wahid et al¹⁴³ used a CNN model based on the 3D residual U-net architecture to segment oropharyngeal cancers based on 5 multiparametric MRI inputs (T2-weighted and T1-weighted image, ADC map, as well as volume transfer constant [K_trans] and extravascular extracellular volume fraction [Ve] derived from DCE MRI). The findings demonstrated that combinations of these inputs enabled better delineation of the cancer tissue than T2-weighted images alone. Segmentation of glioblastoma using deep learning has been typically performed with multiparametric MRI, including all T1-weighted, contrast-enhanced T1-weighted, T2-weighted, and FLAIR images.^144–146 However, in clinical examinations, 1 or more of these scan types is often not usable, for example, because of patient motion. Conte et al¹⁴⁷ developed 2 generative adversarial networks to generate T1-weighted and FLAIR images from contrast-enhanced T1-weighted and T2-weighted images, respectively, and successfully segmented glioblastoma using deep learning only with contrast-enhanced T1-weighted and T2-weighted images. In line with this study, Thomas et al¹³⁹ synthesized either missing T1-weighted, contrast-enhanced T1-weighted, T2-weighted, or FLAIR images from available sequences, using a generative adversarial network model with a tumor-targeting loss to improve the synthesis of tumor areas. Across tumor areas and missing sequences, their approach outperformed conventional approaches for handling missing data (entering a blank mask or copying an existing sequence), particularly when FLAIR images were missing. However, no method could reliably replace missing contrast-enhanced T1-weighted images.

Multiparametric MRI has also been used for the classification or diagnosis of diseases using deep learning.^{135,150–158} Hagiwara et al¹³⁵ applied CNN to T1, T2, and proton density maps of the brain to extract features common to multiple sclerosis and neuromyelitis optica spectrum disorder, another demyelinating disease, and used this information to differentiate these 2 disorders, achieving an accuracy of 80%. Some studies showed the superiority of the diagnostic ability of CNN over radiomics based on multiparametric MRI.^155–158 For example, Truhn et al¹⁵⁵ developed both CNN and radiomic models to differentiate benign from malignant breast lesions using multiparametric MRI data of T2-weighted images, as well as precontrast and 4 postcontrast DCE T1-weighted images, with the CNN showing a significantly higher AUC than the radiomic model. Zhou et al¹⁵⁷ focused on nonmass breast lesions using multiparametric MRI (3 DCE maps related to wash-in, maximum, and wash-out) and also showed better performance with their CNN model than the radiomic model in differentiating between benign and malignant lesions. Furthermore, Zhu et al¹⁵⁹ compared the performances of 1 multi-input and 2 single-input models based on DWI and DCE MRI in differentiating benign from malignant breast lesions and showed that the combination of DCE with DWI was superior to using a single sequence. Previous studies combined the segmentation and diagnosis of tumors in organs such as the prostate,¹³³ breast,¹⁵⁹ and brain^160–162 using deep learning based on multiparametric MRI. Such “1-stop” deep learning algorithms may be widely used in the near future as a computer-aided diagnosis tool to assist clinicians and radiologists by speeding up segmentation, training residents, and providing a preliminary diagnosis.

One of the limitations of deep learning is the lack of transparency associated with how a deep learning model comes to a decision (the so-called “black box problem”), which hinders the trust and acceptance of deep learning in the medical field. Hence, explainable AI models are needed that quantify why certain predictions were made.¹⁶³ For this purpose, several methods have been proposed to visualize the image regions that are important for the AI to make a decision. One of the most popular algorithms is the gradient-weighted class activation mapping (Grad-CAM) method, which generates heat maps to help understand which parts of an input image are important for a classification decision.¹⁶⁴ Grad-CAM has been applied in deep learning studies, including those using multiparametric MRI.^165–169 Grad-CAM was applied to automatic prostate tumor segmentation models to interpret the segmentation results, and Grad-CAM heat maps could differentiate between tumor and normal prostate tissue, indicating that the image information in the tumor was crucial for segmentation by the CNN algorithms.^165,169 Zhang et al.¹⁶⁶ used Grad-CAM to interpret the CNN models trained to classify multiple sclerosis subtypes based on multiparametric MRI data of the brain. The 95th percentile values of Grad-CAM in secondary progressive multiple sclerosis were significantly higher than those in relapsing-remitting multiple sclerosis, suggesting greater heterogeneity. Furthermore, a voxel-wise statistical comparison of thresholded Grad-CAM confirmed the difference between relapsing-remitting multiple sclerosis and secondary progressive multiple sclerosis in discriminative brain regions (Fig. 6).

FIGURE 6: 3D visualization of Grad-CAM for the relapsing-remitting multiple sclerosis and secondary progressive multiple sclerosis group differences. Results from 2-sample t test, which identify the most critical areas of the brain that distinguish the 2 groups (top row: secondary progressive multiple sclerosis > relapsing-remitting multiple sclerosis; bottom row: relapsing-remitting multiple sclerosis > secondary progressive multiple sclerosis). Shown are the t values (plus 95% confidence interval) that demonstrate significance in voxel-wise statistics. The higher the t values, the more significant. Reproduced with permission from Zhang et al.¹⁶⁶

Another limitation of deep learning is its dependency on large datasets. This can be reduced by techniques such as data augmentation and transfer learning. Transfer learning reuses a model already developed for 1 task as the starting point for a model applied to another task. This is especially advantageous for medical tasks because it not only obviates the need for enormous datasets but is less computationally expensive.¹⁷⁰ Another approach is increasing the number of datasets by distributing the deep learning algorithms for training on a large body of decentralized data without directly exchanging the actual data.¹⁷¹ This approach is called federated learning, which mitigates the fundamental problems of patient privacy and data ownership. Federated learning potentially speeds up the development and validation of deep learning algorithms. Federated learning has been utilized in segmenting brain tumors using multiparametric MRI.^172,173 However, the potential trade-off between performance and decentralization (ie, not gathering the data for centralized learning) is unknown.

GENERATING NEW QUANTITATIVE MAPS FROM MULTIPARAMETRIC MRI DATA

The MRI signal is dependent on a multitudinous number of subvoxel tissue properties, which potentially enables the detection of changes occurring at a scale much smaller than the image resolution. However, conventional MRI only provides an indirect measure of tissue properties. Multiparametric MRI may overcome this limitation by exploiting the differing degrees of their contribution to MRI contrasts between physiological and pathological substrates. We can combine multiparametric MRI to create a new metric reflecting some meaningful biological information. There are data-driven and model-driven methods that combine multiparametric MRI for this purpose.¹⁷⁴

Data-Driven Method

The data-driven method uses multivariate analysis tools, machine learning models, or both that can differentiate between overlapping and complementary information of multiparametric MRI.¹⁷⁴ An example is a multivariate myelin model that aimed to generate a cortical myelin map using magnetization transfer ratio, T2*, cortical thickness, and B0 orientation maps.¹⁷⁵ Magnetization transfer ratio and T2* are both sensitive to myelin content. However, the magnetization transfer ratio is also affected by inflammation and tissue pH, and T2* is affected by iron content and tissue orientation. Mangeat et al¹⁷⁵ combined these maps using independent component analysis to identify their shared information, namely, myelin, while correcting for partial volume effect and fiber orientation using a multivariate model. Ciccarelli et al¹⁷⁶ measured N-acetylaspartate levels by MR spectroscopy and axial diffusivity by DWI in the spinal cord of patients with multiple sclerosis and then input these parameters with the cross-sectional area into a statistical model to estimate impairment of mitochondrial metabolism. Lower residual variance of N-acetylaspartate, supposed to reflect impaired mitochondrial metabolism, was associated with higher clinical disability independent of structural damage.

Model-Based Method

This approach is based on biological models, aiming to combine parameters using biophysical or signal models and create a metric that is more accurate or specific than the original metrics.¹⁷⁴ An example is g-ratio, the ratio of the inner to the outer diameter of a nerve fiber, which is associated with the speed of conduction along the axon.¹⁷⁷ Stikov et al¹⁷⁸ combined myelin volume fraction, which can be estimated by techniques such as magnetization transfer imaging¹⁷⁹ and relaxometry,²⁹ with DWI sensitive to axonal volume fraction, such as neurite orientation dispersion and density index, or neurite orientation dispersion and density imaging,¹⁸⁰ to estimate MR g-ratio. A neurite is any process extending from the cell body of a neuron. The parameter g-ratio may enable the differentiation between demyelination and axonal degeneration (Fig. 7).^181–185 Demyelination and remyelination increase and decrease g-ratio, respectively.

FIGURE 7: Representative images from a patient with multiple sclerosis. Synthetic T2-weighted image (A) and maps of myelin volume fraction (B), axon volume fraction (C), and g-ratio (D) are shown. Two plaques are designated by arrows in these images. Even though myelin is severely damaged in these plaques (B, 5.53% and 7.23%), the degrees of axon damage are milder (C, 31.30% and 22.95%). Because myelin is severely damaged in these plaques, corresponding g-ratios are close to 1.00 (D, 0.94 and 0.91). Reproduced with permission from Hagiwara et al.¹⁸¹

Because g-ratio is associated with fiber conduction, the combination of g-ratio and connectome analysis has the potential for investigating brain structure and function. In healthy individuals, the g-ratio distribution across the edges of the graph did not show the power-law distribution observed with a conventional method using the number of streamlines as a weight.¹⁸⁶ Furthermore, significant differences in g-ratio were observed regarding hub structures, suggesting that connections involving hub regions present higher myelination than peripheral connections. Using g-ratio-based connectomes, Kamagata et al¹⁸⁷ examined patients with multiple sclerosis. A g-ratio-based graph theory analysis detected significantly higher nodal strength in patients with multiple sclerosis than in healthy participants. In contrast, the conventional connectome based on the number of streamlines failed to detect any difference. Significant areas were predominantly localized to the limbic area.

Another model-based metric derived from multiparametric MRI is the aerobic glycolytic index (AGI), which captures the metabolic status of tumor tissues.¹⁸⁸ Glycolysis is often augmented in malignant cells, even in the presence of abundant oxygen, in a process known as aerobic glycolysis or Warburg effect.¹⁸⁹ This abnormal metabolism leads to a remarkable decrease in local extracellular pH as a result of increased production of lactic acid, which contributes to tumor malignancy.¹⁹⁰ Assumed to reflect the degree of aerobic glycolysis, AGI can be calculated with amine CEST spin- and gradient-echo echoplanar imaging (CEST-SAGE-EPI)^191–193 and DSC perfusion imaging. Amine CEST quantifies pH-dependent chemical exchange between amine protons and bulk water as magnetization ratio asymmetry. Multiecho readout of SAGE-EPI can be used for calculating the reversible transverse relaxation rate R2', which is sensitive to tissue hypoxia. Aerobic glycolytic index can be calculated by taking the ratio of tumor acidity measured as magnetization ratio asymmetry to relative oxygen extraction fraction, which is calculated by normalizing R2' by CBV. Aerobic glycolytic index was higher in IDH wild-type and high-grade gliomas compared with IDH mutant and low-grade gliomas, respectively.¹⁸⁸ Furthermore, AGI showed a strong correlation with ¹⁸F-fluorodeoxyglucose (¹⁸F-FDG) uptake (Fig. 8). High and low ¹⁸F-FDG uptakes corresponded well with high and low AGI values, respectively. ¹⁸F-FDG PET shows high physiological uptake in the cortex and can obscure ¹⁸F-FDG uptake by a tumor, whereas AGI does not show high values in the cortex. Thus, AGI is potentially more useful than ¹⁸F-FDG PET in clinical settings.

FIGURE 8: Patients with diffuse gliomas with high (A), medium (B), and low (C) AGI and high (A) and low (B) ¹⁸F-FDG uptake. IDH wild-type glioblastomas (A and B) show higher AGI than IDH mutant grade II glioma (C). The contrast-enhancing region of the tumor shows higher AGI than the non-contrast-enhancing region of the tumor (A and B). MTR_asym at 3.0 ppm is grossly high in the entire tumors for both glioblastomas, whereas AGI seems to be correlated more with ¹⁸F-FDG PET than MTR_asym at 3.0 ppm. Whereas ¹⁸F-FDG PET shows high physiological uptake in the cortex, quantitative maps, including AGI, derived from MRI do not show high values in the cortex. NAWM indicates normal-appearing white matter. Reproduced with permission from Hagiwara et al.¹⁸⁸

There are some limitations in combining multiparametric MRI based on separate acquisitions to create new quantitative maps. For example, DWI is typically obtained with single-shot echo-planar imaging that can be affected by susceptibility-induced geometric distortions.¹⁹⁴ Such distortions should be corrected before linear registration with other images to achieve accurate registration and correct resulting maps. Furthermore, any bias in each sequence to be used will propagate into the final results. Besides systematic biases, MRI is prone to image noise. In a simple example of adding 2 maps to create a new map, the variance of the new map σ is:

$\sqrt{{\sigma }_1^2+{\sigma }_2^2},$

where σ₁ and σ₂ are the variances of the 2 original maps.

CONCLUSIONS

Multiparametric MRI has the potential to provide complementary information and overcome the limitations of individual techniques. With the development of rapid acquisition techniques, including simultaneous multiparametric acquisition, the number of parameters that can be acquired in a single clinical scan session has been increasing. Many methods can be used to analyze multiparametric MRI data as a whole rather than each parameter separately. They include clinical scoring systems, machine learning, radiomics, and deep learning. Scoring systems require the input of experts and are still affected by intrarater and interrater variability. Machine learning, radiomics, and deep learning enable automatic image analysis but are still almost exclusively used in research settings because of a lack of generalizability and reproducibility. However, simultaneous multiparametric acquisition techniques may provide reproducible results by such analysis methods because of their quantitative nature, and they have the potential to serve as endpoints in future clinical trials. Other techniques combine multiparametric MRI to create new quantitative maps that capture some meaningful aspects of human biology. They include the MR g-ratio, the inner to the outer diameter of a nerve fiber, which is associated with the speed of conduction along the nerve fiber, and the AGI, which reflects the metabolic status of tumor tissues.