Pilot Study: Texture Analysis of PET Imaging Demonstrates Changes in ¹⁸F-FDG Uptake of the Brain After Prophylactic Cranial Irradiation

Patient Selection

This retrospective study was approved by the institutional Human Subjects Protection Program with a waiver of the requirement to obtain informed consent. Patients were selected from a dataset previously generated by Eshghi et al. (9). In brief, those authors searched the institutional electronic medical record for patients with biopsy-proven small cell lung cancer who were treated with PCI between 2013 and 2016. Patients were treated to a total dose of 2,500 cGy in 10 daily fractions of 250 cGy each, typically delivered over 2 wk. The target volume consisted of the entire brain and surrounding meninges within the cranium, with beam flashing beyond the skin and with the bottom of the field either at the foramen magnum or inferior to C1. Standard blocking of organs at risk such as the lenses, pharynx, and oral cavity was achieved with multileaf collimation. All patients had previously been treated with first-line chemotherapy and thoracic radiation per the standard of care. Patients with brain metastases or other intracranial pathology were excluded. These patients received ¹⁸F-FDG PET/CT imaging before and after PCI therapy as standard-of-care imaging for postchemoradiation tumor response. In total, 16 patients matched the inclusion criteria for this study. A single patient was excluded from the analysis because of poor image quality.

PET/CT Imaging

¹⁸F-FDG PET/CT imaging was performed in accordance with the standard institutional protocol. Patients fasted for a minimum of 4 h, and fingerstick blood glucose levels were measured before administration of ¹⁸F-FDG. ¹⁸F-FDG was administered intravenously at a weight-based dose of 3.7 MBq (0.1 mCi)/kg, with a range of 185–370 MBq (5–10 mCi), after which the patients sat quietly awake for approximately 60 min. Imaging was then performed from the vertex to the thighs using a GE Healthcare 690 time-of-flight PET/CT scanner. A low-dose CT scan was obtained with oral contrast medium but no intravenous contrast medium before the PET acquisition. PET data were acquired using 7–8 bed positions, with 2.5 min per bed position, and were reconstructed using ordered-subsets expectation maximization (28 subsets, 2 iterations).

Image Analysis

The attenuation-corrected, time-of-flight PET image sequences were selected for use in this study. Slices including the brain from the vertex to the brain stem were manually selected. Each slice was segmented manually using ImageJ software in order to select only the patient’s head and to exclude the background and the patient’s arms. These segmented data were used to perform texture analysis of the entire brain parenchyma. Then, the image sets were segmented manually to isolate the right and left frontal, parietal, and temporal lobes. The occipital lobes were excluded from this analysis. Manual segmentation was performed by a radiology resident with prior image segmentation experience. Texture analysis was performed within each brain region on the right and the left.

Texture Analysis

The texture features were computed by first constructing and analyzing the gray-level cooccurrence matrix (GLCM). The GLCM was chosen over other texture analysis methodologies because of its widespread use and validation in the literature and its ability to quantify pixel-scale variations in image brightness. Additionally, given that a manual segmentation method was used, having a robust feature set that is insensitive to individual pixel values was of interest to reduce the overall effect of any variability that may occur due to the manual segmentation procedure.

The GLCM is a spatial histogram that describes the distribution of gray-level values occurring near one another in an image (10). Each entry in the GLCM, p(i,j | d,θ), corresponds to the probability that a pixel (or voxel) with a gray level of i is a distance d pixels away in the θ direction from a pixel with a gray-level of j. With image data quantized into N_g gray levels, the GLCM is an N_g × N_g matrix. In this case, 4 directions for θ are possible: 0°, 45°, 90°, and 135°; for 3-dimensional data, 13 directions are possible for θ. In this study, d was fixed at 1 pixel, and the GLCM was computed for the 4 possible directions of θ for a single image slice. All images were normalized and quantized to 8-bit (intensity ranging between 0 and 255). From the GLCM, we then computed 13 texture features introduced by Haralick et al. (10), averaged over the 13 directions for θ. Statistical significance for differences in texture features was computed using a 2-way dependent Student t test. The analyses were completed using the Python programming language on a computer with an Intel Core I-4710HQ central processing unit (2.50 GHz) and 16-GB double-data-rate 3L memory.