Comparison of intensity normalization methods in prostate, brain, and breast cancer multi-parametric magnetic resonance imaging

doi:10.21203/rs.3.rs-4289787/v1

Intensity variation in multi-parametric magnetic resonance imaging (MP-MRI) poses challenges for MRI analyses. Previous studies have employed several normalization methods, but a lack of consensus on which results in the most comparable images exists. This study analyzed MP-MRI data from patients with prostate, brain, or breast cancer to evaluate common intensity normalization methods to best harmonize intensities across sites, vendors, and magnetic field strengths. Additionally, we calculated radiomic features after normalization to assess downstream effects. In prostate cancer, normalization methods were tested on T2-weighted imaging (T2WI), including a subset scanned with and without an endorectal coil (ERC). In glioblastoma (GBM) patients, methods were evaluated on various MRI sequences. For breast cancer, methods were tested on T1-weighted nonfat-suppressed images. All methods were compared using a two one-sided test (TOST) to test for equivalence of mean and standard deviation of intensity distributions. Across every comparison, the Z-score of intensity within a mask of the organ consistently provided an equivalent distribution (all p < 0.001). Likewise, Z-score normalization provided the highest percentage of radiomic features that were statistically. Our results suggest that normalizing intensity using the Z-score of MRI intensity within an organ mask produces the most comparable intensities across multiple settings.

Biological sciences/Cancer/Breast cancer

Biological sciences/Cancer/Cancer imaging

Biological sciences/Cancer/Cns cancer

Biological sciences/Cancer/Urological cancer/Prostate cancer

Physical sciences/Mathematics and computing/Computational science

MRI

prostate cancer

brain cancer

breast cancer

normalization

radiomics

Multi-parametric magnetic resonance imaging (MP-MRI) is used to assess cancer and response to therapy. Specific to prostate cancer, a typical MP-MRI protocol contains T2-weighted (T2W), diffusion-weighted (DWI), and dynamic contrast enhanced (DCE) imaging. The Prostate and Breast Imaging Reporting and Data Systems, PI-RADS and BI-RADS, respectively, assign a score to MR images and have standardized acquisition, interpretation, and reporting of prostate and breast MRI, as well as aid in the accurate detection of cancerous lesions¹. Moreover, MP-MRI including T1-weighted imaging pre- and post-gadolinium contrast agent (T1 and T1C, respectively) is used to maximize the efficiency of surgical resection and radiation treatment, as well as monitoring progression, for glioblastoma.

While MP-MRI acquisitions are well established techniques for imaging several organs, voxel intensities in “weighted” scans are nonquantitative and can vary within and across patients, tissues, and MRI vendors. Clinically, the most used MRI acquisitions include pre- and post-contrast T1-weighted, T2-weighted, and diffusion weighted imaging (DWI). These scans are assessed qualitatively to determine cancer presence, although apparent diffusion coefficient maps (ADC) can be created from DWI for quantitative assessment. Acquisitions including MR fingerprinting (MRF), advanced diffusion, and a variety of quantitative MRI (QMRI) have been an area of interest for both response assessments in clinical trials and multi-institutional studies. These acquisitions however are not used clinically due to long scan times and variability in acquisition parameters and post-processing techniques^2-4. To make inter- and intra-patient quantitative comparisons, such as with radiomic analyses, images need to be intensity normalized as a pre-processing step. Furthermore, normalization is necessary for the development of MRI-based machine learning techniques for diagnosis of cancer. There is no current gold standard method for signal intensity normalization. Prior studies have normalized by average voxel values within fat and muscle tissue regions^5-7, used N4 bias field correction and intensity Z-score^8-10, and histogram matching and mapping techniques¹¹ to normalize images. Tissue-based normalization has shown to improve inter-patient intensity differences better than unnormalized data and histogram-based normalization methods¹². While Z-score mapping is common among MRI analyses for several disease states^13-16, it can be confounded by factors such as tumor volume and aggressiveness (i.e., increased hypointensity). Additionally, histogram matching and mapping techniques have been shown to be beneficial in normalizing brain MRI¹⁷; however, histogram matching was performed after fat, bone, and background removal, indicating that global normalization of other abdominal organs may be less successful.

Diffusion weighted imaging measures the diffusion of water molecules to generate contrast in MR images. DWI has been shown to detect cancerous tumors and evaluate tumor aggressiveness^4,18,19, but much like T1 and T2WI, DWI is also assessed qualitatively by radiologists. Calculation of ADC from multiple b-values allows a quantitative assessment of water diffusion. Previous studies have shown that ADC has an inverse relationship with higher risk prostate, brain, and breast cancers^20-23. While ADC is considered quantitative, factors such as perfusion can affect lower b-values. Previous studies have assessed normalizing ADC maps prior to analysis. One such study found that a signal-to-noise (SNR)-weighted regularization of ADC produced homogenous maps at varying levels of SNR compared to non-regularized maps which could only estimate ADC accurately at high SNR levels²⁴. Conversely, a study comparing normalizing ADC by the ratio of non-enhancing tumor to normal white matter in high-grade glioma patients showed that normalization did not improve ADC correlations with overall survival²⁵.

Though the need for intensity normalization is well understood, the lack of normalization standards makes it difficult to compare MRI-based analyses. This study analyzed a variety of imaging acquisitions across multiple organs to determine if a universal normalization method could be applied. Specifically, we assessed T2WI collected from prostate cancer patients; T1, T1C, fluid-attenuated inversion recovery (FLAIR), and ADC images collected from GBM patients; and T1-weighted nonfat-suppressed images (T1nFS) from breast cancer patients across three unique sites, multiple clinical MR vendors, and 1.5T and 3T magnetic field strength to examine commonly used post-acquisition intensity normalization methods to identify which method produces images most comparable across vendors for each tissue. Additionally, we examined T2WI collected from prostate cancer patients with an endorectal coil in place and following ERC removal to determine which normalization method best compares these images. Furthermore, we calculated 218 radiomic features across all images to determine how radiomic features are affected by each normalization method. Overall, we tested the hypothesis that normalizing images using signal intensities within a defined region would produce intensity distributions that are most comparable across sites, MRI vendors, and magnetic field strength than unnormalized data.

2.1. Prostate Cancer Cohort

From our TOST test results, we found that prostate Site 1 v Site 2 mean intensity using the standard deviation within the prostate mask as well as all site comparisons using the Z-score normalization were significantly equivalent (all p < 0.001) Similarly, the standard deviation normalization was significantly similar between GE v Siemens, and Z-score normalization was significantly equivalent between all vendors (all p < 0.001). Finally, Z-score normalization was significantly equivalent between 3 T v 1.5 T magnetic field strength as well as between ERC v nERC (both p < 0.001) (Table 1, Figure 1).

Table 1: Mean and standard deviation of intensity for the five prostate normalization methods across each intensity comparison. Abbreviations: ERC = endorectal coil; nERC = post-endorectal coil removal.

Normalization Method	Comparison	Pooled St. Deviation	p-value
Mean Intensity
Unnormalized	Site 1 v Site 2	1	1
	Site 1 v Site 3	1.08E+03	1
	Site 2 v Site 3	95.78	1
Standard Deviation	Site 1 v Site 2	0.50	1
	Site 1 v Site 3	0.48	< 0.001
	Site 2 v Site 3	0.38	1
Z-Score	Site 1 v Site 2	0.01	< 0.001
	Site 1 v Site 3	0.01	< 0.001
	Site 2 v Site 3	0.00	< 0.001
Bladder ROI	Site 1 v Site 2	0.44	0.80
	Site 1 v Site 3	0.47	0.16
	Site 2 v Site 3	0.38	< 0.001
Muscle ROI	Site 1 v Site 2	4.35	1
	Site 1 v Site 3	4.02	1
	Site 2 v Site 3	0.35	0.58
Unnormalized	GE v Siemens	893.91	1
	GE v Philips	1.11E+03	1
	Philips v Siemens	239.94	1
Standard Deviation	GE v Siemens	0.48	< 0.001
	GE v Philips	0.53	1
	Philips v Siemens	0.44	1
Z-Score	GE v Siemens	0.01	< 0.001
	GE v Philips	1.49E-09	< 0.001
	Philips v Siemens	0.01	< 0.001
Bladder ROI	GE v Siemens	0.45	0.99
	GE v Philips	0.46	1
	Philips v Siemens	0.32	< 0.001
Muscle Mean	GE v Siemens	3.86	1
	GE v Philips	4.85	1
	Philips v Siemens	0.53	0.92
Unnormalized	3 T v 1.5 T	1.14E+03	1
Standard Deviation	3 T v 1.5 T	0.48	1
Z-Score	3 T v 1.5 T	0.01	< 0.001
Bladder ROI	3 T v 1.5 T	0.46	0.24
Muscle ROI	3 T v 1.5 T	3.85	1
Unnormalized	ERC v nERC	796.02	1
Standard Deviation	ERC v nERC	0.47	1
Z-Score	ERC v nERC	0.02	< 0.001
Bladder ROI	ERC v nERC	0.47	0.61
Muscle ROI	ERC v nERC	3.74	1
Standard Deviation of Intensity
Unnormalized	Site 1 v Site 3	2.39E+05	1
	Site 1 v Site 2	2.20E+05	1
	Site 2 v Site 3	2.20E+04	1
Standard Deviation	Site 1 v Site 3	111.32	1
	Site 1 v Site 2	107.09	1
	Site 2 v Site 3	87.33	0.65
Z-Score	Site 1 v Site 3	111.32	1
	Site 1 v Site 2	107.09	1
	Site 2 v Site 3	87.33	0.65
Bladder ROI	Site 1 v Site 3	95.05	1
	Site 1 v Site 2	89.15	1
	Site 2 v Site 3	38.04	0.94
Muscle ROI	Site 1 v Site 3	853.40	1
	Site 1 v Site 2	786.42	1
	Site 2 v Site 3	46.44	1
Unnormalized	GE v Siemens	1.91E+05	1
	GE v Philips	2.39E+05	1
	Philips v Siemens	4.09E+04	1
Standard Deviation	GE v Siemens	113.10	1
	GE v Philips	111.51	1
	Philips v Siemens	108.91	1
Z-Score	GE v Siemens	113.10	1
	GE v Philips	111.51	1
	Philips v Siemens	108.91	1
Bladder ROI	GE v Siemens	83.95	1
	GE v Philips	101.10	1
	Philips v Siemens	38.52	0.58
Muscle Mean	GE v Siemens	757.20	1
	GE v Philips	952.62	1
	Philips v Siemens	85.08	1
Unnormalized	3 T v 1.5 T	2.29E+05	1
Standard Deviation	3 T v 1.5 T	121.18	1
Z-Score	3 T v 1.5 T	121.18	1
Bladder ROI	3 T v 1.5 T	89.74	1
Muscle ROI	3 T v 1.5 T	754.55	1
Unnormalized	ERC v nERC	1.70E+05	1
Standard Deviation	ERC v nERC	89.68	1
Z-Score	ERC v nERC	89.68	1
Bladder ROI	ERC v nERC	80.81	1
Muscle ROI	ERC v nERC	754.79	1

2.2. Glioblastoma Cohort

Across T1 images, we found that Site 3 v Site 2 mean intensity using the standard deviation within the brain mask and the mean intensity within the CSF mask, as well as all site comparisons using the Z-score normalization and were significantly equivalent (all p < 0.001) Similarly, the standard deviation, Z-score, and CSF normalization methods were significantly similar between GE v Siemens (all p < 0.001). Finally, Z-score and CSF normalizations were significantly equivalent between 3 T v 1.5 T magnetic field strength (both p < 0.001) (Table 2, Figure 2-4 A).

Table 2: Mean and standard deviation for each of the five intensity normalization methods applied to T1 brain imaging. Abbreviation: CSF = cerebral spinal fluid.

Normalization Method	Comparison	Pooled St. Deviation	p-value
Mean Intensity
Unnormalized	Site 1 v Site 2	115.41	1
	Site 1 v Site 3	217.32	1
	Site 2 v Site 3	150.48	1
Standard Deviation	Site 1 v Site 2	0.10	0.54
	Site 1 v Site 3	0.15	< 0.001
	Site 2 v Site 3	0.11	< 0.001
Z-Score	Site 1 v Site 2	0.00	< 0.001
	Site 1 v Site 3	0.00	< 0.001
	Site 2 v Site 3	0.00	< 0.001
CSF Mask	Site 1 v Site 2	0.03	< 0.001
	Site 1 v Site 3	0.03	< 0.001
	Site 2 v Site 3	0.03	< 0.001
Scaled	Site 1 v Site 2	4.32	0.69
	Site 1 v Site 3	3.76	1
	Site 2 v Site 3	4.02	1
Unnormalized	GE v Siemens	163.54	1
Standard Deviation	GE v Siemens	0.12	< 0.001
Z-Score	GE v Siemens	0.00	< 0.001
CSF Mask	GE v Siemens	0.04	< 0.001
Scaled	GE v Siemens	4.26	1
Unnormalized	3 T v 1.5 T	233.70	0.58
Standard Deviation	3 T v 1.5 T	0.17	< 0.001
Z-Score	3 T v 1.5 T	0.00	< 0.001
CSF Mask	3 T v 1.5 T	0.05	< 0.001
Scaled	3 T v 1.5 T	5.51	0.91
Standard Deviation of Intensity
Unnormalized	Site 1 v Site 2	212.04	0.52
	Site 1 v Site 3	437.71	1
	Site 2 v Site 3	296.73	1
Standard Deviation	Site 1 v Site 2	0.16	0.02
	Site 1 v Site 3	0.28	< 0.001
	Site 2 v Site 3	0.22	0.003
Z-Score	Site 1 v Site 2	0.04	< 0.001
	Site 1 v Site 3	0.02	< 0.001
	Site 2 v Site 3	0.03	< 0.001
CSF Mask	Site 1 v Site 2	0.05	< 0.001
	Site 1 v Site 3	0.07	< 0.001
	Site 2 v Site 3	0.05	< 0.001
Scaled	Site 1 v Site 2	6.08	1
	Site 1 v Site 3	6.67	1
	Site 2 v Site 3	6.28	1
Unnormalized	GE v Siemens	328.81	1
Standard Deviation	GE v Siemens	0.22	0.001
Z-Score	GE v Siemens	0.03	< 0.001
CSF Mask	GE v Siemens	0.06	< 0.001
Scaled	GE v Siemens	8.05	1
Unnormalized	3 T v 1.5 T	401.41	1
Standard Deviation	3 T v 1.5 T	0.25	< 0.001
Z-Score	3 T v 1.5 T	0.05	< 0.001
CSF Mask	3 T v 1.5 T	0.08	< 0.001
Scaled	3 T v 1.5 T	7.88	1

Across T1C images, we found that Site 3 v Site 2 mean intensity using the standard deviation within the brain mask, as well as all site comparisons using the Z-score and CSF normalizations were significantly equivalent (all p < 0.001) Similarly, the Z-score and CSF normalization methods were significantly similar between GE v Siemens (both p < 0.001). Finally, standard deviation, Z-score, and CSF normalizations were significantly equivalent between 3 T v 1.5 T magnetic field strength (all p < 0.001) (Table 3, Figure 2-4 B).

Table 3: Mean and standard deviation for each of the five intensity normalization methods applied to T1C brain imaging. Abbreviation: CSF = cerebral spinal fluid.

Normalization Method	Comparison	Pooled St. Deviation	p-value
Mean Intensity
Unnormalized	Site 1 v Site 2	108.98	1
	Site 1 v Site 3	125.69	0.98
	Site 2 v Site 3	100.67	1
Standard Deviation	Site 1 v Site 2	0.10	0.93
	Site 1 v Site 3	0.11	< 0.001
	Site 2 v Site 3	0.09	< 0.001
Z-Score	Site 1 v Site 2	0.00	< 0.001
	Site 1 v Site 3	0.00	< 0.001
	Site 2 v Site 3	0.00	< 0.001
CSF Mask	Site 1 v Site 2	0.03	< 0.001
	Site 1 v Site 3	0.03	< 0.001
	Site 2 v Site 3	0.03	< 0.001
Scaled	Site 1 v Site 2	3.42	1
	Site 1 v Site 3	3.22	1
	Site 2 v Site 3	3.18	1
Unnormalized	GE v Siemens	114.07	1
Standard Deviation	GE v Siemens	0.11	< 0.001
Z-Score	GE v Siemens	0.00	< 0.001
CSF Mask	GE v Siemens	0.03	< 0.001
Scaled	GE v Siemens	3.37	1
Unnormalized	3 T v 1.5 T	150.40	1
Standard Deviation	3 T v 1.5 T	0.13	< 0.001
Z-Score	3 T v 1.5 T	0.00	< 0.001
CSF Mask	3 T v 1.5 T	0.05	< 0.001
Scaled	3 T v 1.5 T	3.43	1
Standard Deviation of Intensity
Unnormalized	Site 1 v Site 2	208.00	1
	Site 1 v Site 3	230.36	0.89
	Site 2 v Site 3	187.15	1
Standard Deviation	Site 1 v Site 2	0.19	0.95
	Site 1 v Site 3	0.18	0.03
	Site 2 v Site 3	0.17	< 0.001
Z-Score	Site 1 v Site 2	0.04	< 0.001
	Site 1 v Site 3	0.02	< 0.001
	Site 2 v Site 3	0.03	< 0.001
CSF Mask	Site 1 v Site 2	0.04	< 0.001
	Site 1 v Site 3	0.07	< 0.001
	Site 2 v Site 3	0.05	< 0.001
Scaled	Site 1 v Site 2	5.37	0.66
	Site 1 v Site 3	5.72	0.98
	Site 2 v Site 3	5.52	1
Unnormalized	GE v Siemens	208.05	1
Standard Deviation	GE v Siemens	0.19	< 0.001
Z-Score	GE v Siemens	0.03	< 0.001
CSF Mask	GE v Siemens	0.05	< 0.001
Scaled	GE v Siemens	5.54	1
Unnormalized	3 T v 1.5 T	235.99	0.96
Standard Deviation	3 T v 1.5 T	0.19	0.001
Z-Score	3 T v 1.5 T	0.05	< 0.001
CSF Mask	3 T v 1.5 T	0.07	< 0.001
Scaled	3 T v 1.5 T	5.63	1

Across FLAIR images, we found that Site 1 v Site 2 and Site 3 v Site 2 mean intensity using the standard deviation within the brain mask, as well as all site comparisons using the Z-score normalization and were significantly equivalent (p = 0.02 and < 0.001; all p < 0.001). The standard deviation and Z-score normalization methods were significantly similar between GE v Siemens (both p < 0.001). As with T1C magnetic field strength comparisons, standard deviation, Z-score, and CSF normalizations were significantly equivalent between 3 T v 1.5 T (all p < 0.001) (Table 4, Figure 2-4 C).

Table 4: Mean and standard deviation for each of the five intensity normalization methods applied to FLAIR brain imaging. Abbreviation: CSF = cerebral spinal fluid.

Normalization Method	Comparison	Pooled St. Deviation	p-value
Mean Intensity
Unnormalized	Site 1 v Site 2	73.32	1
	Site 1 v Site 3	115.81	1
	Site 2 v Site 3	49.84	1
Standard Deviation	Site 1 v Site 2	0.10	0.10
	Site 1 v Site 3	0.07	< 0.001
	Site 2 v Site 3	0.08	< 0.001
Z-Score	Site 1 v Site 2	0.00	< 0.001
	Site 1 v Site 3	0.00	< 0.001
	Site 2 v Site 3	0.00	< 0.001
CSF Mask	Site 1 v Site 2	0.04	< 0.001
	Site 1 v Site 3	0.11	< 0.001
	Site 2 v Site 3	0.07	< 0.001
Scaled	Site 1 v Site 2	3.68	1
	Site 1 v Site 3	3.53	0.98
	Site 2 v Site 3	3.38	1
Unnormalized	GE v Siemens	86.61	1
Standard Deviation	GE v Siemens	0.09	< 0.001
Z-Score	GE v Siemens	0.00	< 0.001
CSF Mask	GE v Siemens	0.08	< 0.001
Scaled	GE v Siemens	3.63	1
Unnormalized	3 T v 1.5 T	253.28	0.98
Standard Deviation	3 T v 1.5 T	0.11	< 0.001
Z-Score	3 T v 1.5 T	0.00	< 0.001
CSF Mask	3 T v 1.5 T	0.10	< 0.001
Scaled	3 T v 1.5 T	5.54	1
Standard Deviation of Intensity
Unnormalized	Site 1 v Site 2	147.28	1
	Site 1 v Site 3	214.78	1
	Site 2 v Site 3	81.62	1
Standard Deviation	Site 1 v Site 2	0.18	0.90
	Site 1 v Site 3	0.11	< 0.001
	Site 2 v Site 3	0.15	< 0.001
Z-Score	Site 1 v Site 2	0.04	< 0.001
	Site 1 v Site 3	0.02	< 0.001
	Site 2 v Site 3	0.03	< 0.001
CSF Mask	Site 1 v Site 2	0.07	1
	Site 1 v Site 3	0.27	< 0.001
	Site 2 v Site 3	0.18	< 0.001
Scaled	Site 1 v Site 2	6.74	1
	Site 1 v Site 3	6.44	1
	Site 2 v Site 3	6.39	1
Unnormalized	GE v Siemens	176.61	1
Standard Deviation	GE v Siemens	0.16	< 0.001
Z-Score	GE v Siemens	0.03	< 0.001
CSF Mask	GE v Siemens	0.19	< 0.001
Scaled	GE v Siemens	6.57	1
Unnormalized	3 T v 1.5 T	570.24	0.92
Standard Deviation	3 T v 1.5 T	0.18	0.01
Z-Score	3 T v 1.5 T	0.05	< 0.001
CSF Mask	3 T v 1.5 T	0.21	< 0.001
Scaled	3 T v 1.5 T	9.16	1

Finally, across ADC images, we found that Site 3 v Site 2 mean intensity using the standard deviation within the brain mask, as well as all site comparisons using the Z-score normalization and were significantly equivalent (all p < 0.001). As with FLAIR vendor normalizations, the standard deviation and Z-score normalization methods were significantly similar between GE v Siemens (both p < 0.001). Finally, the standard deviation and Z-score normalization methods were significantly similar between 1.5 T v 3 T magnetic field strength (both p < 0.001) (Table 5, Figure 2-4 D).

Table 5: Mean and standard deviation for each of the five intensity normalization methods applied to ADC brain imaging. Abbreviation: CSF = cerebral spinal fluid.

Normalization Method	Comparison	Pooled St. Deviation	p-value
Mean Intensity
Unnormalized	Site 1 v Site 2	54.22	1
	Site 1 v Site 3	1.16E+08	1
	Site 2 v Site 3	7.73E+07	1
Standard Deviation	Site 1 v Site 2	67.02	0.79
	Site 1 v Site 3	28.97	1
	Site 2 v Site 3	51.17	1
Z-Score	Site 1 v Site 2	67.07	0.79
	Site 1 v Site 3	28.98	1
	Site 2 v Site 3	51.21	1
CSF Mask	Site 1 v Site 2	0.03	< 0.001
	Site 1 v Site 3	0.05	< 0.001
	Site 2 v Site 3	0.04	< 0.001
Scaled	Site 1 v Site 2	3.33	1
	Site 1 v Site 3	3.41	1
	Site 2 v Site 3	2.91	1
Unnormalized	GE v Siemens	7.71E+07	1
Standard Deviation	GE v Siemens	53.45	1
Z-Score	GE v Siemens	53.49	1
CSF Mask	GE v Siemens	0.04	< 0.001
Scaled	GE v Siemens	3.85	1
Unnormalized	3 T v 1.5 T	8.69E+07	1
Standard Deviation	3 T v 1.5 T	54.49	0.80
Z-Score	3 T v 1.5 T	54.55	0.80
CSF Mask	3 T v 1.5 T	0.04	< 0.001
Scaled	3 T v 1.5 T	3.53	1
Standard Deviation of Intensity
Unnormalized	Site 1 v Site 2	117.01	1
	Site 1 v Site 3	2.85E+08	1
	Site 2 v Site 3	1.89E+08	1
Standard Deviation	Site 1 v Site 2	188.56	0.91
	Site 1 v Site 3	61.30	1
	Site 2 v Site 3	148.32	1
Z-Score	Site 1 v Site 2	188.64	0.91
	Site 1 v Site 3	61.31	1
	Site 2 v Site 3	148.38	1
CSF Mask	Site 1 v Site 2	0.04	< 0.001
	Site 1 v Site 3	0.16	< 0.001
	Site 2 v Site 3	0.11	< 0.001
Scaled	Site 1 v Site 2	4.57	1
	Site 1 v Site 3	5.44	1
	Site 2 v Site 3	4.68	1
Unnormalized	GE v Siemens	1.89E+08	1
Standard Deviation	GE v Siemens	150.25	1
Z-Score	GE v Siemens	150.32	1
CSF Mask	GE v Siemens	0.11	< 0.001
Scaled	GE v Siemens	5.56	0.85
Unnormalized	3 T v 1.5 T	2.14E+08	1
Standard Deviation	3 T v 1.5 T	153.58	0.89
Z-Score	3 T v 1.5 T	153.65	0.89
CSF Mask	3 T v 1.5 T	0.12	< 0.001
Scaled	3 T v 1.5 T	5.05	1

2.3. Breast Cancer Cohort

In the breast imaging, Z-score normalization was significantly equivalent between all sites mean intensities, as well as the standard deviation of the Z-scored intensities between Site 1 and Site 3 (all p < 0.001). Interestingly, the standard deviation of the raw intensities between Site 2 and Site 3 were significantly equivalent (p < 0.001) (Table 6, Figure 5).

Table 6: Mean and standard deviation of each of the five intensity normalization methods applied to breast imaging.

Normalization Method	Comparison	Pooled St. Deviation	p-value
Mean Intensity
Unnormalized	Site 1 v Site 2	500.89	1
	Site 1 v Site 3	290.65	0.32
	Site 2 v Site 3	547.86	1
Standard Deviation	Site 1 v Site 2	0.74	1
	Site 1 v Site 3	0.58	0.77
	Site 2 v Site 3	0.75	1
Z-Score	Site 1 v Site 2	9.95E+10	< 0.001
	Site 1 v Site 3	7.83E+10	< 0.001
	Site 2 v Site 3	1.11E+09	< 0.001
Sternum Mask	Site 1 v Site 2	2.80	1
	Site 1 v Site 3	4.04	0.86
	Site 2 v Site 3	2.62	1
Thorax Mask	Site 1 v Site 2	17.06	1
	Site 1 v Site 3	22.30	1
	Site 2 v Site 3	13.03	1
Unnormalized	GE v Siemens	179.33	1
Standard Deviation	GE v Siemens	166.01	0.56
Z-Score	GE v Siemens	147.54	1
Sternum Mask	GE v Siemens	179.33	1
Thorax Mask	GE v Siemens	1.60E+03	1
Unnormalized	3 T v 1.5 T	1.73E+05	1
Standard Deviation	3 T v 1.5 T	157.34	1
Z-Score	3 T v 1.5 T	157.34	1
Sternum Mask	3 T v 1.5 T	2.03E+03	1
Thorax Mask	3 T v 1.5 T	1.29E+04	1
Standard Deviation of Intensity
Unnormalized	Site 1 v Site 2	480.05	1
	Site 1 v Site 3	0.87	0.53
	Site 2 v Site 3	9.83E-10	< 0.001
Standard Deviation	Site 1 v Site 2	3.42	0.99
	Site 1 v Site 3	22.04	1
	Site 2 v Site 3	468.82	1
Z-Score	Site 1 v Site 2	0.87	0.74
	Site 1 v Site 3	9.83E-10	< 0.001
	Site 2 v Site 3	3.41	0.99
Sternum Mask	Site 1 v Site 2	21.79	1
	Site 1 v Site 3	1.78E+05	0.75
	Site 2 v Site 3	1.57E+05	0.90
Thorax Mask	Site 1 v Site 2	2.13E+05	0.69
	Site 1 v Site 3	166.01	0.56
	Site 2 v Site 3	147.54	1
Unnormalized	GE v Siemens	2.40E+03	0.98
Standard Deviation	GE v Siemens	1.51E+03	1
Z-Score	GE v Siemens	9.82E+03	1
Sternum Mask	GE v Siemens	1.33E+04	1
Thorax Mask	GE v Siemens	7.25E+03	1
Unnormalized	3 T v 1.5 T	1.83E+05	1
Standard Deviation	3 T v 1.5 T	169.74	0.64
Z-Score	3 T v 1.5 T	169.74	0.64
Sternum Mask	3 T v 1.5 T	2.04E+03	1
Thorax Mask	3 T v 1.5 T	1.28E+04	1

2.4. Radiomic feature analysis

Similarly to the general intensity analysis, each organ and acquisition had unique results; however, there were general trends across all analyses (Figure 6). Z-score normalization had the highest number of features that were statistically equal across all acquisitions. Local Intensity, Intensity Based Statistics, Intensity Histogram, and overall Intensity (i.e., all intensity features) had the lowest number of statistically equal features, apart from Intensity Volume Histogram which remained relatively constant at roughly 50% of features being statistically equal. GLCM had a higher number of statistically equal features across standard deviation and z-score normalization for all organs and acquisitions, with a high number also in CSF Mask brain normalization. TOST results for each organ can be found in Supplemental Tables 1-6. As may be visualized in Figure 2, the Site 3 ADC images were not initially scaled consistently with values ranging from millions to 10^-6. Radiomic features were calculated on images scaled to match units. ADC also had the fewest stable radiomic features across every comparison.

In this study, MP-MRI intensity distributions were assessed to determine the best MR image intensity normalization method for use with quantitative analyses in prostate, glioblastoma, and breast cancer imaging. Two one-sided (TOST) test was used to compare MRI intensities across sites, vendors, and magnetic field strengths used in the three organs, as well using an endorectal coil in prostate imaging. Endorectal coil usage has begun transitioning out of the clinical standard^26-28, thus datasets containing both images with and without an ERC may be impacted by signal intensity differences. Our results suggest that the best normalization for each image acquisition varies; however, in each tested organ and acquisition, the Z-score normalization method produced comparable images across site, vendors, magnetic field strength, and ERC usage. This can be observed visually using the distributions plots and corresponding maps. Likewise, in our radiomic feature analyses, Z-scored normalization had the highest percentage of statistically equal features. These results indicate that a Z-scored normalization could be applied universally across tissue types.

The standard deviation or Z-score of intensity within each organ was expected to have been skewed due to tumor heterogeneity, including tumor volume and aggressiveness, across patients unrelated to MR vendor differences; however, our results found that normalization using these methods, particularly Z-score, produced the most consistent intensities across vendors and endorectal coil usage. Conversely, ROI-based normalization should have addressed the issue of tumor heterogeneity by using intensities external to the organ; however, we found that ROI-based normalization methods performed poorly in comparison to whole-tissue-based normalization. We also expected the thorax masked breast normalization to perform best among the breast normalization methods, however, it is worthwhile to note that signal heterogeneity exists across breast MR images and few options to test masks external to the breast itself are available.

Intensity normalization is imperative to reduce MRI heterogeneity for quantitative analyses across patients and institutions. While many MRI intensity normalization methods have been established, there is no gold standard method to use, further challenging inter-institutional comparisons. One previous study compared the impact of four normalization methods across T2WI before and after radical external beam radiotherapy (RT) on downstream radiomic feature computations²⁹. Their methods included (1) unnormalized images, (2) a centered Z-score using mean and standard deviation of image intensity (i.e., Z-score + 3 times the standard deviation), (3) the centered Z-score using the mean and standard deviation of intensity within the bladder, and (4) a histogram-matching approach as proposed by³⁰. They found that both normalization using the centered Z-score of the image intensity and histogram matching provided the most reproducible radiomic features, whereas ROI-based normalization performed poorly.

In this study, we tested commonly used normalization methods on T2WI across sites, vendors, magnetic field strength, and T2WI across patients scanned with and without an endorectal coil in prostate cancer imaging; T1 non-fat saturated imaging by vendor for breast cancer MRI; and T1, T1C, FLAIR, and ADC in glioblastoma patient imaging to determine the method that produces intensity distributions most similar. Of the methods tested across each tissue type, we found that using Z-scored normalization produces similar intensity distributions across all comparisons, vendors, magnetic field strength, and images with and without an ERC. We additionally calculated 218 radiomic features across images from all normalization methods and found that Z-scored normalization had the highest number of stable features across each comparison. These findings suggest normalization methodology plays a critical role in making inter- and intra-patient MP-MRI-based comparisons.

While the results of this study have promising results, there are several limitations to be considered. One limitation of this study is the relatively small patient cohort compared to previous MP-MRI analyses for both the prostate and glioblastoma cohorts. Additionally, only two MR vendors were compared across images for the glioblastoma and breast, and there were also significantly fewer patients scanned on the Philips scanner for the prostate patients. The lack of patients scanned using the Philips scanner could result in intensity distributions that are less representative than a larger patient cohort may see. Likewise, we did not find any statistically equal radiomic features across neither the GE v Philips nor Siemens v Philips analyses which may also be due to the low population of patients imaged using the Philips scanner. Furthermore, using clinical imaging acquisitions for the glioblastoma and breast cancer cohorts caused a lack of standardized acquisition parameters which may have differing results when controlling for factors such as field strength. Future studies should compare other clinical MR vendors and imaging sequences to increase generalizability of normalization methods. Likewise, imaging phantoms using several MR vendors may provide a more precise estimate of intensity distributions after normalization, as tissue properties between patients is a confounding factor in this study. Finally, we only calculated mean and standard deviation as evaluation metrics for the normalization methods. Future studies should compare additional evaluation metrics to make precise estimates on the most comprehensive normalization metric.

We demonstrate in a cohort of 641 prostate cancer patients, 68 of which had scans with and without the use an endorectal coil, 1016 glioblastoma patients, and 236 female breast cancer patients, that a Z-scored intensity normalization provides distributions that are most comparable across sites, MR vendors, magnetic field strength, prostate ERC usage, and radiomic feature stability. Using a normalization method that best distributes intensity across tissues could help improve quantitative assessments of cancer MRI.Future studies should investigate larger populations as well as additional MR vendors to determine how normalization methods affect downstream analyses of multi-parametric MR images.

Data from three unique sites per organ (prostate, glioblastoma, and breast) were assessed for this study. Details from each site are further detailed in the subsequent sections; however, a simplified table of these data sites and organs is provided in Table 7.

Table 7. Breakdown of prostate, glioblastoma, and breast cancer data by data site, MR manufacturer, and magnetic field strength.

		Demographics		MR Vendor			Magnetic Field Strength
		Patients	Sex	GE	Siemens	Philips	1.5 T	3T
Prostate	Total	641	M: 641	256	295	90	89	552
	Site 1	385	M: 385	256	125	4	3	382
	Site 2	86	M: 86	0	0	0	86	0
	Site 3	170	M: 170	0	170	86	0	170
Glioblastoma	Total	1016	M: 615	410	606	0	51	965
	Total	1016	F: 401	410	606	0	51	965
	Site 1	52	M:35	36	16	-	36	16
	Site 1	52	F: 17	36	16	-	36	16
	Site 2	590	M: 358	0	590	-	15	575
	Site 2	590	F: 232	0	590	-	15	575
	Site 3	374	M: 222	374	0	-	0	374
	Site 3	374	F: 152	374	0	-	0	374
Breast	Total	236	F: 236	190	46	0	185	51
	Site 1	68	F: 68	68	0	-	68	0
	Site 2	100	F: 100	54	46	-	49	51
	Site 3	68	F: 68	68	0	-	68	0

4.1. Prostate Cancer Cohort

4.1.1. Site 1 – LOCAL

Data from 385 prospectively recruited patients treated locally at our institution (Table 7; Figure 7 A, top) with pathologically confirmed prostate cancer undergoing radical prostatectomy between 2014 and 2023 were analyzed for this institutional review board (IRB) approved study. Written informed consent was obtained from all patients. Inclusion criteria for this cohort included clinical imaging including T2-weighted imaging prior to surgery.

Patients underwent multi-parametric magnetic resonance imaging (MP-MRI) prior to prostatectomy on 1.5 T (n_1.5T = 3) or 3T (n_3T = 382) GE (n_GE = 256), Siemens (n_S= 125) or Philips (n_P = 4) MRI scanner (General Electric, Waukesha, WI, USA; Siemens Healthineers, Erlangen, Germany; Philips, Amsterdam, Netherlands) (Figure 7, B). A subset of patients (n = 68) had additional imaging after removal of the endorectal coil on either the GE or Siemens scanner (n_GE = 52, n_S = 16) (Figure 7, C). Each protocol included T2-weighted imaging with acquisition parameters as follows: repetition time (TR) = 3370 milliseconds, FOV = 120 mm, voxel dimensions = 0.23 × 0.23 × 3 mm, acquisition matrix = 512, and slices = 26. All image contrasts used in this study were acquired axially.

4.1.2. Site 2 – PROSTATE-DIAGNOSIS

A publicly available dataset including prostate T2WI scanned on a 1.5 T Philips Achieva using a combined surface and endorectal coil was used for our second site^31,32. From a total of 92 patients, images from 86 patients were ultimately used in this analysis due to image quality (Table 7; Figure 7 A, middle).

4.1.3. Site 3 – PROSTATEx

The final dataset used in this analysis was a collection of retrospective prostate MR studies including T2WI acquired on two different 3T Siemens MR scanners (MAGNETOM Trio and Skyra)^32,33. T2W imaging acquisition parameters include a turbo spin echo sequence with a resolution of ~0.5 mm in plane and a slice thickness of 3.6 mm. All images were acquired without an endorectal coil. After exclusion of images with poor quality, a total of 170 patients’ images were used (Table 7; Figure 7 A, bottom).

4.2. Glioblastoma Cohort

4.2.1. Site 1 – LOCAL

Written, informed consent was obtained from 52 patients for this cohort, each diagnosed with a glioblastoma in concordance with the 2021 WHO classification standards for brain tumors. Inclusion criteria for this cohort included autopsy confirmed GBM and axial clinical imaging including pre- and post-contrast T1-weighted images (T1, T1C), FLAIR, and DWI 1.5 T (n_1.5T = 36, n_3T = 16, n_GE = 36, n_S= 16). Due to the use of clinical imaging, acquisition parameters were not standardized across patients. Axial T1, T1C, FLAIR, and ADC images were selected as the primary acquisitions for this study. ADC maps were calculated using the patient’s clinical DWI. T1, T1C, and ADC images were rigidly aligned to patient’s FLAIR image using SPM12 (https://www.fil.ion.ucl.ac.uk/spm/software/spm12/) (Table 7; Figure 8 A, B, C, D top rows). Examples of images scanned on the GE and Siemens scanners in Figure 8 are from this dataset.

4.2.2. Site 2 – UPENN-GBM

Data from this online repository includes MP-MRI for de novo GBM patients from the University of Pennsylvania Health System^32,34. All axial images in this dataset, including T1, T1C, FLAIR, and ADC, were skull-stripped co-registered by an automated computational method¹⁰. A total of 590 patients from this dataset were used after excluding images without all four pre-surgery acquisitions or poor quality (Table 7; Figure 8 A, B, C, D middle rows).

4.2.3. Site 3 – UCSF-PDGM

Site 3 data comes from the publicly available University of California San Francisco Preoperative Diffuse Glioma MRI (UCSF-PDGM) dataset^32,35. This dataset includes 501 subjects with histopathologically-proven diffuse gliomas who were imaged with a preoperative MRI using a 3T GE Discovery 750. Each image contrast was registered to the FLAIR image (1 mm isotropic resolution) using automated non-linear registration (Advanced Normalization Tools). Resampled co-registered data were then skull stripped using a publicly available deep-learning algorithm^36,37 Table 7; Figure 8 A, B, C, D bottom rows). Though a total of 501 adult patients with pathologically confirmed grade II-IV diffuse gliomas were collected for this database, only the 374 patients with confirmed GBM were used.

4.3. Breast Cancer Cohort

All datasets used for our breast imaging analyses were available online (https://cancerimagingarchive.net)³² and analysis was performed on non-fat suppressed T1 images (T1nFS) (Figure 9).

4.3.1. Site 1 – ACRIN 6698

The ACRIN trial 6698, organized by the American College of Radiology Imaging Network, was a multi-institutional research project^38,39. Its purpose was to determine the efficacy of quantitative DWI in measuring the response of breast cancer to neoadjuvant chemotherapy (NAC). A total 406 women with invasive breast cancer were prospectively enrolled to ACRIN 6698 at ten institutions between August 2012 to January 2015. However, after applying our exclusion criteria described previously in 2.3. Breast Cancer Cohort, only 68 patients’ images were assessed. All patients underwent breast MRI at 4 timepoints over the course of NAC, though only the pre-treatment images are analyzed in this study. MR imaging was performed on a 1.5T GE scanner using a dedicated breast radiofrequency coil (Figure 9 A, top; Figure 9 top). Detailed MRI protocol parameter specifications can be found on https://cancerimagingarchive.net/⁴⁰.

4.3.2. Site 2 – Duke-Breast-Cancer-MRI

This breast cancer cohort was downloaded from the publicly available MRI dataset ⁴¹. The Duke-Breast-Cancer-MRI dataset contains 922 female patients recruited between 2000 and 2014, however, only 351 patients were included in our analyses due to availability of T1nFS images and image quality. Because of annotation constraints described below, a random selection of 100 patients were chosen from the eligible patients for this analysis. As with our local GBM cohort, clinical imaging was provided in the dataset, thus acquisition parameters were not standardized across patients (n_1.5T = 49, n_3T = 51, n_GE= 54, n_S= 46) (Figure 9 A, middle; Figure 9 B).

4.3.3. Site 3 – ISPY2

I-SPY 2 (Investigation of Serial Studies to Predict Your Therapeutic Response with Imaging And moLecular analysis 2) is an ongoing, multi-center study. Its objective is to swiftly assess the effectiveness of novel treatments for breast cancer within the context of NAC⁴². Adult women diagnosed with locally advanced breast cancer (tumor size ≥2.5 cm) without distant metastasis recruited between 2010 and 2016 were analyzed for this study. Breast MRI data was acquired prospectively at over 22 clinical centers using a standardized image acquisition protocol. Patients underwent 4 MRI exams before and during NAC, though only the first scan was assessed in the current study. This is a comprehensive, highly curated imaging data set with histopathologic outcome that can be used to develop, test, and compare imaging metrics and prediction models for breast cancer response to treatment. A total of 719 patients were included in this dataset, however, only 68 were assessed after applying the exclusion criteria. MR imaging was performed on a 1.5T GE scanner. All required imaging was performed axially with full bilateral coverage⁴³ (Figure 9 A, bottom; Figure 9 B top).

4.4. MRI Normalization

Multiple normalization methods were used for each of the three tissue types. Tissue and regions of interest (ROIs) were defined for each tissue type using AFNI (Analysis of Functional NeuroImages, http://afni.nimh.nih.gov/)⁴⁴. Prostate masks were created on each slice of the patient’s T2-weighted image (T2WI) and additional 10-voxel radius circular ROIs were defined on one slice of the patient’s T2WI within the (1) bladder and (2) levator ani muscle. Corresponding masks were created on the T2WI for patients who had an additional scan done post-endorectal coil removal. Examples of these masks can be viewed in Figure 10, top. The three prostate acquisitions tested in this study were normalized using (1) unnormalized, (2) the standard deviation of intensity within the prostate mask, (3) the Z-score of masked intensity, and the mean intensity within the (4) bladder and (5) levator ani muscle.

Brain imaging masks were segmented using SPM12, defined as the combination of the white and gray matter masks. Cerebral spinal fluid (CSF) masks were created by thresholding the ADC for the high diffusion areas, as this is an indicator of fluid. Finally, intensity within the brain mask was scaled into values of 0-255 as used in⁴⁵. Aligned T1, T1C, FLAIR, and ADC were normalized using (1) unnormalized, (2) standard deviation and (3) Z-score of intensity within the brain mask, and mean intensity within the (4) CSF, and (5) scaled intensity (Figure 10, middle).

Breast masks were manually drawn on MR images using ITK-Snap. Due to the size of each patient’s imaging, only the center 15 slices were annotated. Similarly to the previously mentioned organs, a breast mask was first drawn, encompassing the breast tissue only. A mask of the sternum was drawn on the axial images and location was verified using the sagittal and coronal images. Finally, the thorax was annotated, avoiding any additional tissue. Five normalization methods were applied to the T1nFS images: (1) unnormalized, (2) standard deviation and (3) Z-score of intensity within the breast mask and using the mean intensity within the (4) sternum and (5) thorax masks (Figure 10, bottom).

4.5. Radiomic feature calculation

Radiomic features were calculated across each image using Matlab’s radiomics function which calculates a total of 197 features. These include 136 texture features (i.e., 50 gray level co-occurrence matrix (GLCM), 16 gray level dependence zone matrix (GLDZM), 32 gray level run length matrix (GLRLM), 16 gray level size zone matrix (GLSZM), 17 neighboring gray level dependence matrix (NGLDM), and 5 neighboring gray tone difference matrix (NGTDM)), and 61 intensity features (i.e., 18 Intensity Based Statistics, 23 Intensity Histogram, 18 Intensity Volume Histogram, and 2 Local Intensity).

4.6. Statistical analysis

Mean and standard deviation of MRI intensity as well as radiomic features were calculated across patients following normalization. Intensity distributions were compared across sites, MR vendors, magnetic field strength (i.e., 1.5T v 3T) using a two one-sided (TOST) test, a test of equivalence that is based on the classical t-test⁴⁶. While the TOST test requires both one-sided tests to be statistically significant (i.e., < 0.05), all results described below use the highest p-value for each test.

Acknowledgments: We would like to thank our patients for their participation in this study.

Author Contribution: Savannah R. Duenweg: Conceptualization, Methodology, Software, Validation, Formal analysis, Investigation, Writing – Original Draft, Writing – Review & Editing, Visualization. Samuel A. Bobholz: Conceptualization, Formal analysis, Software, Validation, Writing – Review & Editing. Michael J. Barrett: Software, Validation. Allison K. Lowman: Software, Validation, Writing – Review & Editing. Aleksandra Winiarz: Software, Validation, Writing – Review & Editing. Biprojit Nath: Software, Validation, Writing – Review & Editing. Fitzgerald Kyereme: Software. Stephanie Vincent-Sheldon: Writing – Review & Editing. Peter S. LaViolette: Conceptualization, Methodology, Validation, Formal analysis, Investigation, Funding acquisition, Resources, Data Curation, Writing – Review & Editing, Supervision, Project administration.

Data Availability Statement: Internal data presented in this study are available on request from the corresponding author. The data are not publicly available due to privacy concerns. External data can be found at the citations provided in their respective sections.

Funding: This research was funded by NIH/NCI R01CA218144, R21CA231892, R01CA249882, Advancing a Healthier Wisconsin, Strain for the Brain Inc, the Ryan M. Schaller Foundation, and the State of Wisconsin Tax Check Off Program for Prostate Cancer Research.
Institutional Review Board Statement: The study was conducted according to the guidelines of the Declaration of Helsinki and approved by the Institutional Review Board of MEDICAL COLLEGE OF WISCONSIN (PRO22426, 19 September 2014, PRO26467, 13 July 2017, PRO32772 14 March 2019).

Informed Consent Statement: Informed consent was obtained from all local subjects involved in the study. External cohorts posted online adhered to each local protocol.

Conflicts of Interest: The authors declare that there is no conflict of interest.

Weinreb, J. C. et al. PI-RADS Prostate Imaging - Reporting and Data System: 2015, Version 2. Eur Urol 69, 16-40 (2016). https://doi.org/10.1016/j.eururo.2015.08.052
Gracien, R. M. et al. How stable is quantitative MRI? - Assessment of intra- and inter-scanner-model reproducibility using identical acquisition sequences and data analysis programs. Neuroimage 207, 116364 (2020). https://doi.org/10.1016/j.neuroimage.2019.116364
Keenan, K. E. et al. Recommendations towards standards for quantitative MRI (qMRI) and outstanding needs. J Magn Reson Imaging 49, e26-e39 (2019). https://doi.org/10.1002/jmri.26598
McGarry, S. D. et al. Multi‐Site Concordance of Diffusion‐Weighted Imaging Quantification for Assessing Prostate Cancer Aggressiveness. Journal of Magnetic Resonance Imaging (2021). https://doi.org/10.1002/jmri.27983
Sørland, K. I. et al. Pseudo-T2 mapping for normalization of T2-weighted prostate MRI. MAGMA 35, 573-585 (2022). https://doi.org/10.1007/s10334-022-01003-9
Breto, A. L. et al. Deep Learning Approach for Multi-Reference Tissue Normalization on T2-weighted MRI in Longitudinal Dataset from Prospective Radiotherapy Trial for Prostate Cancer. International Journal of Radiation Oncology*Biology*Physics 108, S130 (2020). https://doi.org/https://doi.org/10.1016/j.ijrobp.2020.07.858
Robitaille, N. et al. Tissue-based MRI intensity standardization: application to multicentric datasets. Int J Biomed Imaging 2012, 347120 (2012). https://doi.org/10.1155/2012/347120
Zheng, H. et al. Integrative Machine Learning Prediction of Prostate Biopsy Results From Negative Multiparametric MRI. J Magn Reson Imaging 55, 100-110 (2022). https://doi.org/10.1002/jmri.27793
Yu, Y. et al. Magnetic resonance imaging radiomics predicts preoperative axillary lymph node metastasis to support surgical decisions and is associated with tumor microenvironment in invasive breast cancer: A machine learning, multicenter study. EBioMedicine 69, 103460 (2021). https://doi.org/10.1016/j.ebiom.2021.103460
Bakas, S. et al. The University of Pennsylvania glioblastoma (UPenn-GBM) cohort: advanced MRI, clinical, genomics, & radiomics. Sci Data 9, 453 (2022). https://doi.org/10.1038/s41597-022-01560-7
Shinohara, R. T. et al. Statistical normalization techniques for magnetic resonance imaging. Neuroimage Clin 6, 9-19 (2014). https://doi.org/10.1016/j.nicl.2014.08.008
Sunoqrot, M. R. S., Nketiah, G. A., Selnæs, K. M., Bathen, T. F. & Elschot, M. Automated reference tissue normalization of T2-weighted MR images of the prostate using object recognition. MAGMA 34, 309-321 (2021). https://doi.org/10.1007/s10334-020-00871-3
Kimura, Y. et al. SISCOM technique with a variable Z score improves detectability of focal cortical dysplasia: a comparative study with MRI. Ann Nucl Med 26, 397-404 (2012). https://doi.org/10.1007/s12149-012-0585-4
Kranzusch, R. et al. Z-score mapping for standardized analysis and reporting of cardiovascular magnetic resonance modified Look-Locker inversion recovery (MOLLI) T1 data: Normal behavior and validation in patients with amyloidosis. J Cardiovasc Magn Reson 22, 6 (2020). https://doi.org/10.1186/s12968-019-0595-7
Liss, M. A. et al. MRI-Derived Restriction Spectrum Imaging Cellularity Index is Associated with High Grade Prostate Cancer on Radical Prostatectomy Specimens. Front Oncol 5, 30 (2015). https://doi.org/10.3389/fonc.2015.00030
Thavendiranathan, P. et al. Myocardial T1 and T2 Mapping by Magnetic Resonance in Patients With Immune Checkpoint Inhibitor-Associated Myocarditis. J Am Coll Cardiol 77, 1503-1516 (2021). https://doi.org/10.1016/j.jacc.2021.01.050
Sun, X. et al. Histogram-based normalization technique on human brain magnetic resonance images from different acquisitions. Biomed Eng Online 14, 73 (2015). https://doi.org/10.1186/s12938-015-0064-y
Woodfield, C. A. et al. Diffusion-weighted MRI of peripheral zone prostate cancer: comparison of tumor apparent diffusion coefficient with Gleason score and percentage of tumor on core biopsy. AJR Am J Roentgenol 194, W316-322 (2010). https://doi.org/10.2214/AJR.09.2651
desouza, N. M., Reinsberg, S. A., Scurr, E. D., Brewster, J. M. & Payne, G. S. Magnetic resonance imaging in prostate cancer: the value of apparent diffusion coefficients for identifying malignant nodules. Br J Radiol 80, 90-95 (2007). https://doi.org/10.1259/bjr/24232319
Duenweg, S. R. et al. Diffusion Restriction Comparison between Gleason 4 Fused Glands and Cribriform Glands within Patient Using Whole-Mount Prostate Pathology as Ground Truth. Tomography 8, 635-643 (2022). https://doi.org/doi:10.3390/tomography8020053
Hurrell, S. L. et al. Optimized b-value selection for the discrimination of prostate cancer grades, including the cribriform pattern, using diffusion weighted imaging. Journal of Medical Imaging 5 (2017). https://doi.org/10.1117/1.jmi.5.1.011004
Darbar, A., Waqas, M., Enam, S. F. & Mahmood, S. D. Use of Preoperative Apparent Diffusion Coefficients to Predict Brain Tumor Grade. Cureus 10, e2284 (2018). https://doi.org/10.7759/cureus.2284
Ren, C., Zou, Y., Zhang, X. & Li, K. Diagnostic value of diffusion-weighted imaging-derived apparent diffusion coefficient and its association with histological prognostic factors in breast cancer. Oncol Lett 18, 3295-3303 (2019). https://doi.org/10.3892/ol.2019.10651
Sveinsson, B., Gold, G. E., Hargreaves, B. A. & Yoon, D. SNR-weighted regularization of ADC estimates from double-echo in steady-state (DESS). Magn Reson Med 81, 711-718 (2019). https://doi.org/10.1002/mrm.27436
Qin, L. et al. Normalization of ADC does not improve correlation with overall survival in patients with high-grade glioma (HGG). J Neurooncol 137, 313-319 (2018). https://doi.org/10.1007/s11060-017-2719-y
Dhatt, R. et al. MRI of the Prostate With and Without Endorectal Coil at 3 T: Correlation With Whole-Mount Histopathologic Gleason Score. AJR Am J Roentgenol 215, 133-141 (2020). https://doi.org/10.2214/AJR.19.22094
Gawlitza, J. et al. Impact of the use of an endorectal coil for 3 T prostate MRI on image quality and cancer detection rate. Sci Rep 7, 40640 (2017). https://doi.org/10.1038/srep40640
Turkbey, B. et al. Comparison of endorectal coil and nonendorectal coil T2W and diffusion-weighted MRI at 3 Tesla for localizing prostate cancer: correlation with whole-mount histopathology. J Magn Reson Imaging 39, 1443-1448 (2014). https://doi.org/10.1002/jmri.24317
Scalco, E. et al. T2w-MRI signal normalization affects radiomics features reproducibility. Med Phys 47, 1680-1691 (2020). https://doi.org/10.1002/mp.14038
Nyúl, L. G., Udupa, J. K. & Zhang, X. New variants of a method of MRI scale standardization. IEEE Trans Med Imaging 19, 143-150 (2000). https://doi.org/10.1109/42.836373
Bloch, B. N., Jain, A. & Jaffe, C. C. (The Cancer Imaging Archive, 2015).
Clark, K. et al. The Cancer Imaging Archive (TCIA): Maintaining and Operating a Public Information Repository. Journal of Digital Imaging 26, 1045-1057 (2013). https://doi.org/https://doi.org/10.1007/s10278-013-9622-7
Litjens, G., Debats, O., Barentsz, J., Karssemeijer, N. & Huisman, H. (The Cancer Imaging Archive, 2017).
Bakas, S. et al. (The Cancer Imaging Archive, 2021).
Calabrese, E. et al. The University of California San Francisco Preoperative Diffuse Glioma MRI Dataset. Radiol Artif Intell 4, e220058 (2022). https://doi.org/10.1148/ryai.220058
Calabrese, E., Villanueva-Meyer, J. E. & Cha, S. A fully automated artificial intelligence method for non-invasive, imaging-based identification of genetic alterations in glioblastomas. Sci Rep 10, 11852 (2020). https://doi.org/10.1038/s41598-020-68857-8
Calabrese, E., Rudie, J. D., Rauschecker, A. M., Villanueva-Meyer, J. E. & Cha, S. Feasibility of Simulated Postcontrast MRI of Glioblastomas and Lower-Grade Gliomas by Using Three-dimensional Fully Convolutional Neural Networks. Radiol Artif Intell 3, e200276 (2021). https://doi.org/10.1148/ryai.2021200276
Newitt, D. C. et al. Test-retest repeatability and reproducibility of ADC measures by breast DWI: Results from the ACRIN 6698 trial. J Magn Reson Imaging 49, 1617-1628 (2019). https://doi.org/10.1002/jmri.26539
Partridge, S. C. et al. Diffusion-weighted MRI Findings Predict Pathologic Response in Neoadjuvant Treatment of Breast Cancer: The ACRIN 6698 Multicenter Trial. Radiology 289, 618-627 (2018). https://doi.org/10.1148/radiol.2018180273
Newitt, D. C. et al. (The Cancer Imaging Archive, 2021).
Saha, A. et al. A machine learning approach to radiogenomics of breast cancer: a study of 922 subjects and 529 DCE-MRI features. Br J Cancer 119, 508-516 (2018). https://doi.org/10.1038/s41416-018-0185-8
Li, W. et al. Predicting breast cancer response to neoadjuvant treatment using multi-feature MRI: results from the I-SPY 2 TRIAL. NPJ Breast Cancer 6, 63 (2020). https://doi.org/10.1038/s41523-020-00203-7
Li, W. et al. (The Cancer Imaging Archive, 2022).
Cox, R. W. AFNI: Software for analysis and visualization of functional magnetic resonance neuroimages. Computers and Biomedical Research 29 (1996). https://doi.org/10.1006/cbmr.1996.0014
Akbari, H. et al. Imaging Surrogates of Infiltration Obtained Via Multiparametric Imaging Pattern Analysis Predict Subsequent Location of Recurrence of Glioblastoma. Neurosurgery 78, 572-580 (2016). https://doi.org/10.1227/NEU.0000000000001202
Lakens, D. Equivalence Tests: A Practical Primer for. Soc Psychol Personal Sci 8, 355-362 (2017). https://doi.org/10.1177/1948550617697177

No competing interests reported.

Comparison of intensity normalization methods in prostate, brain, and breast cancer multi-parametric magnetic resonance imaging

Status:

Version 1

Abstract

Figures

1. Introduction

2. Results

3. Conclusions

4. Materials and Methods

Declarations

References

Additional Declarations

Supplementary Files

Status:

Version 1