John Moeller

Certifying and Removing Disparate Impact

What does it mean for an algorithm to be biased? In U.S. law, unintentional bias is encoded via disparate impact, which occurs when a selection process has widely different outcomes for different groups, even as it appears to be neutral. This legal determination hinges on a definition of a protected class (ethnicity, gender) and an explicit description of the process. When computers are involved, determining disparate impact (and hence bias) is harder. It might not be possible to disclose the process. In addition, even if the process is open, it might be hard to elucidate in a legal setting how the algorithm makes its decisions. Instead of requiring access to the process, we propose making inferences based on the data it uses. We present four contributions. First, we link disparate impact to a measure of classification accuracy that while known, has received relatively little attention. Second, we propose a test for disparate impact based on how well the protected class can be predicted from the other attributes. Third, we describe methods by which data might be made unbiased. Finally, we present empirical evidence supporting the effectiveness of our test for disparate impact and our approach for both masking bias and preserving relevant information in the data. Interestingly, our approach resembles some actual selection practices that have recently received legal scrutiny.

Dynamic field shaping to optimize stereotactic radiosurgery

A dynamic field shaping collimation system is evaluated for use in stereotactic radiosurgery of non-spherical lesions. The concept is as follows: (a) use the existing circular collimators to define a cone which encompasses the maximum dimensions of the target volume; (b) position two sets of independent rectangular photon collimators immediately upstream from the circular aperture and allow each collimator to have independent translational and rotational motion in order to define, for each increment of arc, a polygonal field shape having up to four straight and four curved edges which enscribe the beams eye projection of the target; (c) modify the translational and rotational position of each independent collimator with each change in arc angle to continuously shape the instantaneous field to the target shape. A prototype device has been constructed and uses vane control technology developed in a related research project in electron arc therapy. The efficacy of this device is illustrated by dose calculations and measurement based on actual clinical data. Dose volume histograms are used to compare the dose received by three techniques: single isocenter treatment using a single circular aperture, dual isocenter treatment, and single isocenter treatment using dynamically shaped fields. Doses were calculated throughout the brain using a volume grid of 3 mm spacing. Dose volume histograms comparing dose within the target volume and brain volume excluding target volume, as well as computed isodose distributions, demonstrate the possible reduction in normal tissue dose burden while simultaneously preserving dose uniformity throughout the prescribed target volume. This simple four-vane collimation system may provide a viable alternate treatment technique for non-spherical lesions.

Dynamic wedge field techniques through computer-controlled collimator motion and dose delivery

Clinical treatment planning situations arise which require different wedge angles within segments of a single therapeutic x-ray field. Idealized wedge-shaped dose distributions, including combination of several wedge segments of different angle within a single field, are generated and delivered through computer control of asymmetric collimator motion and dose per field segment. A dual-pass technique is introduced to provide improved adherence to the prescribed isodose distribution. Dynamic wedge distributions are verified by film densitometry and ionization chamber measurement. These results suggest the potential importance of this technique as an added clinical radiotherapy tool.

Facilitation of radiotherapeutic error by computerized record and verify systems

PURPOSE To examine the impact of computerized record and verify (R&V) systems on types of radiotherapeutic error. MATERIALS AND METHODS Radiation therapy treatment errors reported by therapists at the University of Utah between July 1, 1999 and June 30, 2000 were retrospectively reviewed. RESULTS During a 1-year period in which 22,542 external beam radiation therapy treatments were administered, 38 treatment errors (representing 0.17% of external beam treatments administered during this period) were identified and reviewed. Nine cases (0.04% of treatments) representing four types of record and verify (R&V)-related errors were identified, in which the departments R&V system played a contributory role in the treatment error. CONCLUSIONS The common denominator among these R&V-related errors was excessive reliance upon the computer system by therapists. R&V systems eliminate some, but not all, pathways of radiotherapeutic error. Although R&V systems have assumed a crucial role in the precise and reproducible delivery of increasingly complex radiation therapy treatments, their inability to eradicate all radiotherapeutic errors coupled with their parallel ability to facilitate certain mistakes mandates vigilance on the part of the radiation therapy team. Radiation therapy treatment procedures must preserve careful oversight of R&V functions to minimize prospects for treatment error.

High-dose-rate afterloading brachytherapy in carcinoma of the uterine cervix

The Brachytron has been used in the University of California at San Diego Medical Center since 1970 as one method of treating gynecological malignancies. This machine contains a high intensity cobalt 60 remote afterloading cycling source used for intracavitary brachytherapy. One hundred twenty-seven patients with epithelial carcinoma of the cervix are available for analysis of 5-year survival, and 176 are analyzed for treatment complications two years following therapy. Five year survival figures for FIGO-staged patients treated with external beam pelvic irradiation and intracavitary Brachytron treatments are as follows: Stage I, 89%; Stage II, 58%; Stage III, 33%, and two of five patients Stage IVa. Rectal complications graded moderate or severe (M, S) were dose-related and gradually decreased over the years as techniques improved. Complications from early results in 1970-1972 (24% M, 10% S) were reduced to lower levels in 1976-1979 (14% M, 4% S). The Brachytron offers the advantage of rapid dose delivery. Thus, patients can be treated in an outpatient setting, avoiding the cost of hospitalization and the risks of anesthesia. The Brachytron also offers virtually complete radiation safety to all attending medical personnel. With survival and complication figures similar to those reported for patients treated with conventional low-dose-rate brachytherapy, the Brachytron represents an effective alternate mode of therapy for uterine carcinoma.

Dose to the contralateral breast: A comparison of two techniques using the enhanced dynamic wedge versus a standard wedge

The dose to the contralateral breast has been associated with an increased risk of developing a second breast malignancy. Varying techniques have been devised and described in the literature to minimize this dose. Metal beam modifiers such as standard wedges are used to improve the dose distribution in the treated breast, but unfortunately introduce an increased scatter dose outside the treatment field, in particular to the contralateral breast. The enhanced dynamic wedge is a means of remote wedging created by independently moving one collimator jaw through the treatment field during dose delivery. This study is an analysis of differing doses to the contralateral breast using two common clinical set-up techniques with the enhanced dynamic wedge versus the standard metal wedge. A tissue equivalent block (solid water), modeled to represent a typical breast outline, was designed as an insert in a Rando phantom to simulate a standard patient being treated for breast conservation. Tissue equivalent material was then used to complete the natural contour of the breast and to reproduce appropriate build-up and internal scatter. Thermoluminescent dosimeter (TLD) rods were placed at predetermined distances from the geometric beams edge to measure the dose to the contralateral breast. A total of 35 locations were used with five TLDs in each location to verify the accuracy of the measured dose. The radiation techniques used were an isocentric set-up with co-planar, non divergent posterior borders and an isocentric set-up with a half beam block technique utilizing the asymmetric collimator jaw. Each technique used compensating wedges to optimize the dose distribution. A comparison of the dose to the contralateral breast was then made with the enhanced dynamic wedge vs. the standard metal wedge. The measurements revealed a significant reduction in the contralateral breast dose with the enhanced dynamic wedge compared to the standard metal wedge in both set-up techniques. The dose was measured at varying distances from the geometric field edge, ranging from 2 to 8 cm. The average dose with the enhanced dynamic wedge was 2.7-2.8%. The average dose with the standard wedge was 4.0-4.7%. Thermoluminescent dosimeter measurements suggest an increase in both scattered electrons and photons with metal wedges. The enhanced dynamic wedge is a practical clinical advance which improves the dose distribution in patients undergoing breast conservation while at the same time minimizing dose to the contralateral breast, thereby reducing the potential carcinogenic effects.

Electron arc therapy: design, implementation and evaluation of a dynamic multi-vane collimator system.

Innovative techniques in motion control technology have been applied to the design and implementation of a portable computer-controlled multi-vane collimator for use in electron arc therapy. The collimator, consisting of 18 independently controlled vanes, is inserted into the standard accessory mount assembly of a linear accelerator, in the same fashion as standard field shaping blocks. Power is supplied to the collimator vane motors via a self-contained battery system. The range of motion of the vanes, symmetrically mounted nine on each side, provides a variable aperture width projected to isocenter of 2 cm minimum to 8 cm maximum. The projected length of the aperture at isocenter is 38 cm. The transition time between vane positions is less than 1 second, corresponding to gantry movement of less than 1 degree. The movement of each of the 18 vanes is monitored and controlled by six individually addressed three axis processors that are shielded from the electron beam. A table of collimator vane positions versus gantry angle, as determined by dose optimization calculations, is stored in a data file. The desired collimator vane position corresponding to the current arc segment is conveyed from the control console to each vane controller via packets within a token passing network. Communication between the computer in the console area and the vane controllers is accomplished through encoded infra-red pulse transmission, eliminating the need for additional communication lines between the console and the accelerator. This dynamic collimator offers improved dose uniformity while simplifying the delivery of electron arc therapy.

Measurement of mechanical accuracy of isocenter in conventional linear-accelerator-based radiosurgery.

PURPOSE Five Varian linear accelerators were studied to determine whether their mechanical isocentric accuracies were sufficient for radiosurgery and, if not, if the observed errors were sufficiently consistent and predictable to be correctable by some form of secondary collimator steering device to maintain isocentric alignment. METHODS AND MATERIALS A 0.3 mW 670 nm diode laser was mounted in the secondary collimator insert of a radiosurgery extended collimator assembly. A cylindrical lens was used to create a laser fan beam that passed through isocenter and could be oriented parallel or perpendicular to the plane of rotation. A position sensitive photo-diode having an electrical output that varied with the portion of its surface illuminated was mounted at isocenter in a rotational mount. This mount tracked the accelerator gantry such that the surface of the photo-diode remained perpendicular to the laser beam during gantry rotation. An X/Y recorder was connected to the gantry-angle potentiometer of the accelerator and to the photo-diode and plotted the positional variation from isocenter with gantry rotation. RESULTS The root-mean-square error for the five machines was +/- 0.06 to +/- 0.08 mm in the plane of rotation and +/- 0.17 to +/- 0.35 mm out of (perpendicular to) the plane of rotation. The in-plane-of-rotation errors tended to be maximal near the diagonal gantry angles and the out-of-plane-of-rotation errors were maximal in the over and under vertical positions. CONCLUSIONS Both types of errors were predictable but only the out-of-plane-of-rotation errors were considered large enough to warrant consideration of correction (although the need is debatable). On all the tested machines, the out-of-plane-of-rotation error curve was a relatively smooth bell-shaped function that would be readily amenable to correction. The diode laser/photo-detector system used should prove useful in accurately defining isocenter and facilitating the precise adjustment of the laser isocenter lights.

Optimization of electron arc therapy doses by multi-vane collimator control.

Retrospective computer simulations, based on clinical treatment planning data available from over 50 patients treated by electron arc radiotherapy to the chestwall following mastectomy, show that a dramatic improvement in dose uniformity can, in many clinical situations, be achieved by dynamic shaping of the electron arc collimator, under computer control, as a function of gantry angle and distance superior or inferior to the central plane. The greatest improvement in dose uniformity is seen in calculational planes in which the patient contour has the greatest departure from a circular shape. Dosimetric studies demonstrate this improvement. Indicators for use of variable-width multi-vane electron arc collimators include the following: (1) Mechanical constraints of the therapy equipment may limit the placement of isocenter to an inadequate depth which causes large variation in the SSD around the arc; (2) Out of the central plane, the shape of the chest wall may change dramatically across the limits of the arc, creating large variations in the dose distribution; (3) Clinical definition of the treatment surface to include surgical scars or other at-risk volume may create an irregularly shaped treatment surface, thereby changing the fraction of the arc included in the treatment surface from one plane to the next. Electron arc collimator shape determines both the dose rate and the electron arc beam profile. Both the dose rate and the beam profile must be included in the integration of dose to a point within the arc. The dose to a point within the arc can be modified by as much as a factor of 1.5 to 2.0 by increasing the collimator width from 3 cm to 7 cm. A multi-vane collimator allows these changes to be made in each specific plane to compensate for changes in patient contour.

Dosimetric parameters of Enhanced Dynamic Wedge for treatment planning and verification

Treatment planning for Enhanced Dynamic Wedge requires a knowledge of the dosimetric parameters of the treatment fields. These dosimetric parameters include depth doses, surface doses, buildup doses, peripheral doses, beam profiles, wedge angles and wedge factors. These parameters and their application to treatment planning are evaluated and compared with standard open field and metal wedge field dosimetric parameters.