Imaging radiation dose in breast radiotherapy by X-ray CT calibration of Cherenkov light

Abstract

Imaging Cherenkov emission during radiation therapy cancer treatments can provide a real-time, non-contact sampling of the entire dose field. The emitted Cherenkov signal generated is proportional to deposited dose, however, it is affected by attenuation from the intrinsic tissue optical properties of the patient, which in breast, ranges from primarily adipose to fibroglandular tissue. Patients being treated with whole-breast X-ray radiotherapy (n = 13) were imaged for 108 total fractions, to establish correction factors from the linear relationships between Cherenkov light and CT number (HU). This study elucidates this relationship in vivo, and a correction factor approach is used to scale each image to improve the linear correlation between Cherenkov emission intensity and dose ([Formula: see text]). This study provides a major step towards direct quantitative radiation dose imaging in humans by utilizing non-contact camera sensing of Cherenkov emission during the radiation therapy treatment.

Fig. 1. Study setup and patient positioning.

The Cherenkov camera and optical surface guidance projectors and cameras were fixed to the ceiling. The linear accelerator (linac) gantry rotates to each beam position and remains stationary for delivery of each field. When the linac beam of X-rays (yellow) is incident upon the tissue, Cherenkov light is emitted isotropically from within. Some of this light is detected by the Cherenkov camera (blue). The camera intensifier is triggered on during only the linac pulses, thereby suppressing ambient light interference. The optical surface guidance system cameras are used to set up the patient and ensure correct alignment by casting a red, diffuse light pattern onto the patient, and tracking respective movement.

Fig. 2. Model to build correction factors.

Cherenkov frames are recorded throughout a fraction of treatment and summed into a cumulative image (a). Each beam was then separated out by beam energy and gantry angle. Each Cherenkov image is divided by the respective co-registered surface dose image (b) rendered in C-Dose software from the treatment plan (Supplementary Fig. 1). The patient planning CT scan is then sampled down up to 10 mm for an average CT attenuation number (HU) (c), done in the treatment planning software by creating a structure (contour pictured). A correction factor is calculated from both average CT attenuation and Cherenkov normalized by dose. In an ideal scenario, the corrected Cherenkov image corrects for tissue optical properties and shares a higher metric of uniformity with the surface dose image (b).

Fig. 3. Cherenkov image data from 13 whole-breast radiotherapy patients.

Patients 1 (a), 2 (b), 4 (d), 5 (e), 6 (f), 9 (i), 11 (k),  and 12 (l) had prescribed radiation treatments using only 6 MV energy beams (6X), and treatments for Patients 3 (c), 7 (g), 8 (h), 10 (j), and 13 (m) included both 6 and 10 MV beams (6X/10X), indicated in the bottom left corner for each patient thumbnail. The recorded Cherenkov emission is overlain in color on the recorded background image (grayscale), which is captured in real time. In (h), a small patch bolus was used as buildup material over the surgical scar (right lateral, region of higher intensity in the lateral mammary fold). In each thumbnail, Cherenkov fields have undergone spatial and temporal median filtering, and thresholding based on closest match to the treatment field. The images presented correspond to different color scales to optimize dynamic range visibility.

Fig. 4. Variability in Cherenkov intensity with fibroglandular and adipose tissue content.

In a, the CT scan of Patient 3 shows dense, fibroglandular content throughout the breast volume. Ten fractions of Cherenkov imaging are pictured (units photons (γ)), each normalized with respect to dose (Gy) as shown beside this. In (b), the CT scan of Pt 8 is characterized by largely adipose tissue, resulting in a much brighter appearance of Cherenkov images as compared those in a. Subfigures c and d illustrate the linear correlation between HU and the median Cherenkov counts per unit delivered dose (γ Gy⁻¹), averaged over the number of fractions imaged (mean). Beams are separated by color (Entrance/RPO, gray and Exit/LAO, blue). Subfigure (c), n = 13, maps this relationship for all 13 patients treated with 6 MV beams, and (d), n = 5, for the five patients treated with 10X beams. Each point shown is averaged over all fractions for each patient (see Table 1 for number of fractions). RPO and LAO means for each patient were averaged into one point, and the linear correction was computed and applied to each Cherenkov (γ Gy⁻¹) value for both 6X and 10X beams to normalize the Cherenkov signal by HU value (Supplementary Fig. 2b, d). Error bars shown depict the root mean square error.

Fig. 5. Cherenkov images corrected for radiodensity.

The final corrected RPO and LAO images are shown. Subfigure (a) organizes the original (labeled Uncorrected) Cherenkov images recorded during 6 MV radiotherapy treatment (units in Cherenkov photons, γ) as compared to the expected surface dose estimate (labeled Dose) from the treatment plan (Gy). After CT attenuation (HU) corrections have been carried out, the corrected Cherenkov images are shown (labeled Corrected). In (b), the same is shown for 10 MV beams. The corrected Cherenkov images qualitatively match the expected surface dose images more closely than the uncorrected images, where further quantitation is presented in Fig. 6.

Fig. 6. The HU correction increases linearity between Cherenkov light and dose.

The same number of regions of interest (ROI) per patient are sampled from each corrected and uncorrected Cherenkov image stacks and co-registered surface dose images to yield the relationships shown. The median Cherenkov intensities from uncorrected images (a) are plotted against corresponding ROIs from the predicted dose in the treatment plan. The median Cherenkov intensities from the corrected images are then plotted in the adjacent plot (b) for both 6 and 10 MV beams. The y-intercepts have been constrained to cross at the origin, which yields an uncorrected 6X beam regression R²_uncorr,6 = 0.67 (light blue), which exhibits the largest spread of data. The HU-corrected 6X beam regression is improved by 0.18, to R²_corr,6 of 0.85. The linear regression of the 10X beams are improved slightly by 0.04, from 0.91 to 0.95 (dark blue). The Pearson’s linear correlation coefficient increased from r_uncorr,6 = 0.82 uncorrected to r_corr,6 = 0.92 corrected in 6X beams, and from r_uncorr,10 = 0.96 uncorrected to r_corr,10 = 0.97 corrected for the 10X beams. All data are organized by beam energy, where light blue is representative of all 6 MV data and dark blue is representative of all 10 MV data.