Motivation

Whole-body diffusion-weighted imaging (DWI) is a promising approach for detecting prostate-cancer bone metastases.¹ Bone lesions are typically identified qualitatively as hyperintense regions on DWI and can be quantitatively examined by computing the apparent diffusion coefficient (ADC).² However, normal tissues with high signal on DWI can impede lesion detection, and interpretation of ADC values is confounded by numerous physiological factors.³ Multicompartmental modeling of the DWI signal is an effective approach for overcoming these obstacles⁴ that could improve the detection of metastases. In this study, we applied multicompartmental modeling to develop an empirical tissue classifier for identifying bone lesions in whole-body DWI.

Methods

This prospective study included 30 patients with prostate cancer who underwent an extended whole-body MRI examination in addition to routine clinical imaging. Standard-of-care evaluation identified 107 bone lesions in 25 of these patients.

Whole-body MRI acquisition
MR imaging was performed on a 3T clinical scanner (Discovery MR750; GE Healthcare). Five stations were imaged for each patient, corresponding roughly to the head, chest, abdomen, pelvis, and thighs. At each station, an axial volume of multi-shell diffusion data was acquired using 4 b-values: 0, 500, 1000, and 2000 s/mm², sampled at 1, 6, 6, and 12 unique diffusion-encoding gradient directions, respectively (default tensor, TR: 4750ms, TE: 75ms, matrix: 80×80 resampled to 128×128, FOV: 400mm, slices: 46, slice thickness: 6mm). For anatomical reference, a high resolution T2-weighted volume was also acquired at each station with scan-coverage identical to that of the multi-shell DWI volume (TR: 1350ms, TE: 113ms, matrix: 384×224 resampled to 512×512, FOV: 400mm, slices: 46, slice thickness: 6mm).

MRI post-processing
Each multi-shell DWI volume was first corrected for distortions due to B0-inhomogeneity, gradient nonlinearity, and eddy currents.⁵ The signal intensity of each DWI volume was corrected to account for noise.⁶ Isotropic diffusion was assumed, so directional DWI volumes at each b-value were averaged. Conventional ADC maps were computed by fitting the DWI data to a monoexponential signal model.⁷

Regions of interest (ROIs) were defined on DWI volumes over each of the identified bone lesions. In the 5 patients without metastases, control ROIs were defined over the entire body excluding the head.

Multicompartmental modeling
Restriction spectrum imaging (RSI) is a multicompartmental modeling framework for DWI. A previous study (see companion abstract #2625) determined an optimal RSI model for describing whole-body diffusion data: $$S(b)=\sum_{i=1}^{4}C_ie^{-bD_i}$$ where S(b) denotes the measured DWI signal at a particular b-value, C_i describes the compartmental signal contributions to be computed via model-fitting, and D_i refers to the compartmental diffusion coefficients which are fixed for each of the 4 tissue compartments (to 0, 1.2e-3, 2.9e-3, and >3.0e-2mm²/s, respectively). Signal-contribution (C_i) maps were computed for each patient by fitting this model to the signal-vs-b-value curve from each voxel.⁴

Tissue classifier development
Previous studies have shown that diffusion signal in tumors is enriched in the two model compartments with lowest diffusion coefficient, C₁ and C₂.^4,8,9 To determine how C₁ and C₂ vary between cancerous and normal tissue, two joint histograms were computed: one recording the C₁ and C₂ signal-contribution of all voxels within the bone-lesion ROIs, and another from all voxels within the whole-body control ROIs. Normalizing these histograms by the total signal yielded joint C₁,C₂ probability density functions (PDFs) for lesions and normal tissue. Normalizing the lesion PDF by the sum of the lesion and control PDFs generated a posterior-probability distribution for lesion signal, which can serve as a lookup table relating a voxel’s C₁ and C₂ measurements to the likelihood it contains cancerous tissue. This lookup table was used to compute voxel-wise maps of cancer likelihood from the C₁ and C₂ signal-contribution maps of all patients.

Tissue classifier performance
Leave-one-out cross validation at the patient level was used to assess the performance of the RSI-derived tissue classifier. Likelihood values and class labels (positive for voxels within lesion ROIs and negative for the rest of the FOV) were used to generate a receiver operating characteristic (ROC) curve. For comparison, an ROC curve was similarly generated using ADC values. Area under the ROC curve (AUC) and false-positive rate at 80% sensitivity (FPR₈₀) were recorded for both classifiers.

Results

Example RSI signal-contribution maps are shown in Figure 1 for a patient with bone lesions. Figure 2 shows the joint C₁,C₂ PDFs for normal tissue and bone lesions, alongside the lookup table of cancer-likelihood values derived from these PDFs. A voxel-wise cancer likelihood map is shown in Figure 3 for comparison against conventional MR images. Figure 4 illustrates that a tissue classifier based on these likelihood values can more accurately identify metastatic lesions than one based on conventional ADC values.

Discussion

Multicompartmental analysis of diffusion signal from prostate-cancer bone lesions enabled development of an empirical signal classifier for identifying cancerous tissue in DWI data. This classifier relates signal contributions from model compartments with lower diffusion coefficients to the likelihood that such contributions are from cancerous tissue. This approach considerably outperformed a classifier based on conventional ADC values and could potentially help to increase the accuracy of whole-body cancer screening.