High-Precision Radiotherapy and the Computer Vision Project

The aim of the computer vision project is to reduce the patient positioning uncertainty associated with fractionated external beam radiotherapy so that the region of high dose can be designed to conform tightly to the tumor and be delivered with high precision. Theoretically, it then may be possible to escalate the prescribed dose and improve local control without a concomitant increase in normal tissue damage. Components of the computer vision project are being developed and released to the clinic in stages. Ultimately, our aim is to develop an automatic, quick and reliable computer-based system capable of looking into the treatment room to "see" whether or not patients are properly set up for irradiation.

To reduce day-to-day setup variation, a reference image (A) is subtracted from real-time video of the slightly misaligned patient (B). Therapists use the subtraction image to interactively return the patient to the correct position (C).
Figure 1: To reduce day-to-day setup variation, a reference image (A) is subtracted from real-time video of the slightly misaligned patient (B). Therapists use the subtraction image to interactively return the patient to the correct position (C).

In the first clinical release, our aim is to use digital video technology to correct patient setups online, i.e., before irradiation. To accomplish this, video images are obtained and archived on the first day of treatment, after care is taken to accurately position the patient and the attending radiation oncologist approves radiographic images of the setup. The initial set of video images is marked as a reference set and, on subsequent treatment days, the reference images are retrieved from the archive and subtracted in real-time from live video. With the aid of a computer monitor in the treatment room, radiation therapists use the live subtraction images to interactively return the patient to the reference position before radiation is delivered. Since an orthogonal pair of images is used, the online correction is fully 3D. In a controlled study of head and neck patients, the technique was shown to eliminate large setup errors and reduce random setup uncertainty to about 1.5 mm, without a significant increase in overall treatment time or labor-intensive procedures. Currently, the video system is used to position all patients treated with intensity-modulated radiotherapy in our Department. Scott Johnson, Ph.D., leads the development and clinical implementation of the system.

Of course, the ideal treatment position is not necessarily established on the first day of treatment or easily verified by an oncologist using radiographic images. Instead, it is the position at the time of CT simulation that establishes the ideal patient position since it is the CT scan that is used to define the juxtaposition of internal anatomy and create a patient-specific treatment plan. To eliminate CT-to-treatment transfer errors, the second phase of the project focuses on determining exactly how the video cameras view the radiation isocenter of the linear accelerator. In computer space, the results are used to "see" the patient's skin in the CT data from each camera's-eye-view. The rendered images essentially predict how the patient's external anatomy should appear to the cameras when the patient is perfectly aligned for treatment.

For tumors of the abdomen and thorax, careful patient positioning is not sufficient to ensure that the tumor is properly positioned for treatment because of motion due to respiration. The third phase of the computer vision project is aimed at developing a video-based, "no touch" approach to gating radiation treatments for respiration. Using live video feed from a lateral viewing point, we continuously compare the position of the anterior chest wall in the live video with the reference position recorded at the time of CT simulation. A rendered image of the CT data from the lateral viewing perspective is used to define the reference chest wall position. Preliminary work indicates that the technique can be used to detect chest wall displacements as small as 1 mm at frame rates greater than 30 frames/second.

To reduce CT-to-treatment transfer errors, volume-rendered images of the ideal patient position (A) are created so that they have the same viewing perspective of the treatment room as an orthogonal pair of cameras (B). By layering the rendered images on top of live video, therapists can immediately detect and correct positioning errors before radiation is delivered.
Figure 2: To reduce CT-to-treatment transfer errors, volume-rendered images of the ideal patient position (A) are created so that they have the same viewing perspective of the treatment room as an orthogonal pair of cameras (B). By layering the rendered images on top of live video, therapists can immediately detect and correct positioning errors before radiation is delivered.

To gate radiotherapy treatments for respiration, we are developing a no touch technique in which a computer continuously compares the anterior chest wall position seen in live video (A) with the reference position in a rendered image (B).
Figure 3: To gate radiotherapy treatments for respiration, we are developing a no touch technique in which a computer continuously compares the anterior chest wall position seen in live video (A) with the reference position in a rendered image (B).