For adaptive radiotherapy it is common to collect images of the patient throughout the course of therapy. Because of temporal variations, however, it is usually necessary to deform images so as to merge them into a cohesive dataset. This image registration makes the accurate merging of dose distributions difficult, if not impossible. Some have decided to do this by “deforming” the dose distributions, somewhat analogous to deforming the images, but it has been suggested that this is not appropriate. This is the premise debated in this monthˈs Point/Counterpoint. Arguing for the Proposition is Timothy E. Schultheiss, Ph.D. Dr. Schultheiss received his Ph.D. degree in Physics from Brown University in 1979. He has held faculty positions at M.D. Anderson Cancer Center, Fox Chase Cancer Center (Professor and Director of Radiation Physics), and is now Professor and Director of Radiation Physics at the City of Hope Cancer Center. He is a Fellow of the ACR, AAPM, and ASTRO, and is certified in Therapeutic Radiological Physics by the American Board of Radiology. Dr. Schultheiss has served on many AAPM Committees and Task Groups including as Chair of the Statistics and Biological Effects Committees. During his career he has been involved in the premarket deployment of a number of new technologies. He has published extensively in biological effects of radiation, especially radiation myelopathy, also in prostate cancer, statistical analysis of clinical data, and in large-field IMRT. Arguing against the Proposition is Wolfgang A. Tomé, Ph.D. Dr. Tomé obtained his Ph.D. in mathematical physics in 1995 from the University of Florida and completed a post doc and two-year residency in radiation oncology physics at the Shands Cancer Center of the University of Florida in 1998. From 1998 to 2012, he served as faculty member in the Departments of Biomedical Engineering, Human Oncology, and Medical Physics of the University of Wisconsin, where was promoted to Professor with tenure in 2009. He is currently the Director of Physics of the Oncophysics Institute at the Albert Einstein College of Medicine of Yeshiva University and the Director of the Division of Medical Physics of Montefiore Hospital, the teaching hospital of the Albert Einstein College of Medicine. He also holds appointments as Professor of Radiation Oncology at the Albert Einstein College of Medicine and as Visiting Professor of Medical Physics at the Centre of Medical Radiation Physics of the University of Wollongong, Australia. He is board certified by the American Board of Radiology in Therapeutic Radiological Physics and is a Fellow of the AAPM. Dr. Tomé's research interests are bio-mathematical modeling of cancer treatments, biologically guided radiation therapy, adaptive radiation therapy, deformable image registration, 4D patient management, image guided stereotactic body radiotherapy, image guided fractionated stereotactic radiotherapy, and radiosurgery. He has been a member of many AAPM Task Groups and Committees and currently serves on the ASTRO Radiation Oncology Institute Information Technology Infrastructure Committee and the ASTRO Council on Health Policy: Evaluation Subcommittee of the Emerging Technologies Committee. For decades, physicists have striven to increase the accuracy of dose calculation and dose display in radiation therapy treatment planning. Although the leaders in this effort have been amazingly successful, the success of these efforts gives some physicians (and physicists) too much confidence in the dose distributions we see displayed. In fact, we have become far too credulous regarding the barrage of computer output in our field. Although long in coming, the day of deformable image registration (IR) is upon us. The reason it was long in coming is that it is very difficult.1 Each image source will have its own inaccuracies due to distortion, artifacts, resolution, interference, motion, etc. We all have seen rigid IR algorithms find some entirely unexpected (and wrong) solutions. Generally, registration software allows manual adjustment after optimization because the human brain processes gray scale images so much better than a computer.2 However, manual adjustment of the registration is not really possible with deformable registration. With different images, the optimal registration is not objectively definable. That is, there are no objective metrics for assessing the registration one gets when one image is registered and deformed to match another image. Rigid IR alone is fraught with error. Deformable IR is yet more error prone. Now some would add deforming the calculated dose distribution along with the deformed image. With deformed images (or the accompanying contours) we can at least choose to accept or modify them. But when the dose is deformed we have nothing upon which we can base a visual evaluation. Of course, the deforming of dose is well-intentioned.3 The absorbed dose, being a local phenomenon, essentially belongs to the cell that absorbed it. Because cells and tissue move around, we would have to calculate a new dose distribution for each fraction if we want to achieve the greatest accuracy, which is now within our grasp. Then we need some way to merge all of these dose distributions. Enter deformable dose. We can deform all these dose distributions onto a single volumetric study. But let us not be so naïve as to believe we have achieved our goal of ultimate accuracy. The problems are greater than merely those of deformable IR. We are likely to have tissues that simply disappear (along with their dose) during a course of therapy. These include shrinking solid tumors, enlarged lymph nodes, and some normal tissues such as the parotid gland. Some organs may inflate and deflate over time. The lung does this with a period of seconds, but the rectum also does this with a period of hours. The small bowel can slosh about in the abdomen without our being able to tell one loop from another. Finally, tumor growth or inflammation can cause tissues to be present at the end of treatment that were not there at the beginning. The ultimate problem with deformed dose is our inability to measure it. Comparison with measurement is always the standard in the mathematical modeling of physical phenomena. Until we can deform dose with algorithms that have been validated against measurement, rather than being merely based on image manipulation, we should withhold all commercial use of this misleading process. It is more akin to “Photoshopping” the dose than to dose calculation. The goal of adaptive radiation therapy is, first, to determine if the treatment is being delivered as planned, by acquiring additional image sets throughout the course of treatment (in addition to the planning image) and, second, to adjust the treatment plan if objectives are not being met. Without accurately accumulating the dose over multiple images, it could be hazardous to adjust the treatment plan. Consider the following examples. If the paradigm of uniform PTV dose coverage is employed, an adequate approach to determine delivered target dose would be to register the GTV with the planning CT to form a composite GTV and check if this composite GTV lies within the uniform 3D PTV dose distribution. Clearly, this approach does not necessitate dose deformation but only image deformation. However, in the case of dose painting where target dose is heterogeneous, dose warping is necessary to ensure dose to corresponding spatial locations are accurately accumulated.3,4 Not accurately accumulating the dose could be hazardous, as it may lead to treatment decisions being based on incorrect dose distributions. For example, target cold-spots may overlap in reality. Lack of this knowledge could be detrimental to the treatment outcomes, since a significant dose deficit to even a very small portion of a high-risk area within the GTV can have a detrimental effect on the achievable tumor control probability.5 The same also holds for organs at risk, which by treatment plan design see a highly nonuniform dose distribution. Moreover, organ shape, size, and position can change from fraction-to-fraction due to organ motion and filling, and the fact that the treatment target is realigned to correct for possible interfraction target motion.6 Hence, if accurate estimation of expected normal tissue complication probability for organs at risk is desired for plan adaptation then it is necessary to warp the dose. Using image sets acquired just prior to delivery of radiotherapy is, however, only a first order approximation, since things might change during the course of delivery. Before discussing how this point can be addressed, let me just state that the approximation based on a single image set acquired just prior to delivery is still better than assuming that patients are static CT scans and “flying blind.” Ultimately, however, we have to go further: ideally one would acquire anatomical image information and record the machine state and dose delivery status at time points during the delivery. This information could then be used, employing deformable dose accumulation across image sets that are highly correlated, to arrive at a more accurate estimate of dose received for both the target and organs at risk for a given fraction. Dose could then be accumulated over the course of treatment by adding to the record new imaging information along with the dose delivery information from each fraction. The realization of this vision of both dynamically deforming the image and accumulating dose is not too far off into the future and will become clinical reality with the introduction of MR-guided radiation therapy. Professor Tomé has it right. He states that “without accurately accumulating the dose over multiple images, it could be hazardous to adjust the treatment plan.” The problem is that we cannot accurately accumulate dose over multiple images for the reasons stated in my opening remarks. Actually, the problem is worse. There are no metrics upon which to base the accuracy of this dose accumulation. The effort to add doses delivered to the changing anatomy of the patient over a course of treatment is a worthy research endeavor. However, until we are in a position to demonstrate the accuracy of both the deformed image registration and the resultant dose accumulation, such unvalidated software should not be implemented in any clinical setting. We must require that the same rigor be applied to deformed and accumulated dose distributions as is applied to calculated dose distributions. FDA take note. Note that the proposition is stated in the present tense. Only a Luddite would propose that accurately deforming the dose will never be possible or should not be attempted, but it is incumbent upon physicists to ensure that new technologies are deployed with safety as the highest priority. We must be able to test the validity of a deformable image registration—not just that kidney maps to kidney, for example, but that there is voxel-by-voxel agreement. We need measures of the accuracy of the image registration, not just the correlation or mutual information. Like dose accumulation, autocontouring software generally relies upon deformable registration of a target image to a reference image. Extant autocontouring software generally requires significant postprocessing manual corrections or fine tuning of its organ identifications. If we cannot automatically identify entire organs by software, we certainly cannot identify voxels. If we cannot identify and follow voxels, we cannot accurately accumulate dose. I would like to thank my esteemed colleague for the valuable points he has raised in his opening statement and his call to caution. While I agree with Dr. Schultheiss that accurate deformable image registration (DIR) is a tough clinical problem to tackle, it is incumbent on us to educate our physician colleagues on the limitations and applicability of DIR just as our profession has done successfully in the case of advanced modern dose calculation algorithms over the last two decades. Hence, the clinical implementation of DIR should be pursued carefully, deliberately and, in the beginning, only for suitably selected patient groups. Moreover, in my humble opinion, to simply throw in the towel and say that there are too many uncertainties and complexities associated with DIR for it to be clinically useful, is not an acceptable approach. Rather, we should continue in a prospective manner to explore the usefulness of DIR using in silico clinical trials to determine which patient populations and clinical sites could benefit from DIR without actually changing current treatment plans and treatment paradigms. Using this approach will allow one to determine if DIR and deformable dose accumulation (DDA) allows for better prediction of expected normal tissue toxicities. To this end, however, it is essential to have a DDA technique that does not depend on the way dose is accumulated across image sets. Our group has recently presented a DDA methodology that exhibits this property.7 Clearly, DDA and DIR hold great clinical promise for improved prediction of normal tissue toxicities and have the potential to allow one to escalate the dose to target structures at constant expected normal tissue complication probability (NTCP), i.e., allowing one to pursue iso-NTCP escalation strategies. However, this potential application should be carefully explored in prospective clinical trials.