Analysis of heterogeneity in T2-weighted MR images can differentiate pseudoprogression from progression in glioblastoma

Abstract

Purpose: To develop an image analysis technique that distinguishes pseudoprogression from true progression by analyzing tumour heterogeneity in T2-weighted images using topological descriptors of image heterogeneity called Minkowski functionals (MFs).

Methods: Using a retrospective patient cohort (n = 50), and blinded to treatment response outcome, unsupervised feature estimation was performed to investigate MFs for the presence of outliers, potential confounders, and sensitivity to treatment response. The progression and pseudoprogression groups were then unblinded and supervised feature selection was performed using MFs, size and signal intensity features. A support vector machine model was obtained and evaluated using a prospective test cohort.

Results: The model gave a classification accuracy, using a combination of MFs and size features, of more than 85% in both retrospective and prospective datasets. A different feature selection method (Random Forest) and classifier (Lasso) gave the same results. Although not apparent to the reporting radiologist, the T2-weighted hyperintensity phenotype of those patients with progression was heterogeneous, large and frond-like when compared to those with pseudoprogression.

Conclusion: Analysis of heterogeneity, in T2-weighted MR images, which are acquired routinely in the clinic, has the potential to detect an earlier treatment response allowing an early change in treatment strategy. Prospective validation of this technique in larger datasets is required.

Fig 1. Grey-scale thresholding of a region of interest.

A segmented region of interest (ROI) displayed as a binary combination of black and white pixels at 11 different grey-scale thresholds. Each of these black and white images can be characterized by the three 2D MFs; area, perimeter and genus.

Fig 2. MFs of the retrospective patient cohort.

Spectral representations of MFs (mean ± standard error) plotted as a function of grey-scale threshold using blinded (a-c) and unblinded (d-f) data from the retrospective patient cohort. The more heterogeneous the regions of interest, the higher the normalized perimeter value; and the further the genus value is from unity.

Fig 3. Training dataset feature selection.

(a) Graph of -log₁₀P values for all features derived by comparing progression and pseudoprogression datasets (2-tailed unpaired t-test). The significance threshold was set arbitrarily at P < 0.05 (horizontal line) and the selected features are shown as solid black bars. (Abbreviations: nArea, normalized area; nPeri, normalized perimeter; nSI, normalized signal intensity). (b) Heat map showing the selected features from the univariate scaled values of the MFs and size metrics used in the optimal SVM model for the training dataset. Dark blue represents the lowest values and yellow the highest values. Note that patient 20 was the only case of pseudoprogression assigned on clinical grounds in the study and was a false positive. (Abbreviations: PsP, pseudoprogression; P, progression). (c) Bland-Altman plot comparing the difference between two observer measurements, and the mean of the measurements. The bias was 0.4±1.6. The mean bias and the 95% limits of agreement of the interobserver difference are shown as dotted lines.

Fig 4. Training dataset size and signal intensity selected features.

Plot of (a) total perimeter (P = 0.03, t = 2.3, 15 df) and (b) total area (P = 0.02, t = 2.6, 15 df) for patients with progression and pseudoprogression in the training data set. (c) Plot of the normalized minimum signal intensity (P = 0.006, t = 3.2, 15 df) for patients with progression and pseudoprogression in the training data set. (d) Relationship between total area and total perimeter for patients with progression (squares) and pseudoprogression (crosses) in the training data set. Total area is a surrogate metric of volume (a stack of slices summed together) and total perimeter is a surrogate metric of surface area (a stack of slice perimeters summed together). The solid line gives the relationship between volume and surface area for a sphere.

Fig 5. Model comparison with radiation necrosis and vasogenic oedema.

T₂–weighted axial images showing a patient with (a) progression, (b) pseudoprogression, (c) radiation necrosis, and (d) vasogenic oedema. Principal component score plot (e) for those patients with progression (filled black squares) and pseudoprogression (empty squares) from the training dataset, using the features selected for the original SVM model. Vasogenic oedema (triangle) and radiation necrosis (empty circle) were also plotted using the same set of selected features. All features were univariate-scaled. The displayed T2 Hotelling’s tolerance ellipse is set at the 0.05 significance level. The loading plot (f) corresponding to the principal component score plot, showed that the normalized perimeter and size features were positively correlated and separated pseudoprogression from progression and radiation necrosis.