Examining and confronting the challenges of DIR QA in clinical use
Deformable image registration (DIR) continues to be an important tool to improve the quality of treatment planning and patient care in radiation oncology. Yet, many clinicians remain uncertain when and how to use it effectively.
In the simplest terms, deformable image registration is the process of matching two medical images for anatomical alignment. DIR is routinely employed in clinics to assist with diagnostic image registration, contouring, and dose accumulation. Deformed CT, PET, and MR images may provide more specific guidance for treatment planning, re-planning, and re-irradiation cases.
While the clinical uses of DIR vary, the rationalization for deformation is reasonably straightforward. We have a case for deformable registration when patient scans have positional differences, multiple targets, or anatomical changes. For optimal planning, DIR allows clinicians to register two varying patient images with the most accurate anatomical alignment.
Here are a few examples of when deformable registration may be helpful:
While many physicists, dosimetrists, and radiation oncologists recognize the benefit of deformable image registration—as in the cases above—they struggle to assess and trust its accuracy.
We must answer one critical question for effective DIR in clinical use.
Did the registration map the anatomy of interest on my previous scan to the exact same anatomy on my current scan?
This question provides the specificity required to conclusively know if DIR was accurate for the local anatomy we care about during a particular patient’s treatment, providing accuracy and confidence in treatment planning.
Clinicians should assess if the questions they currently ask about the quality of DIR have convenient, rather than conclusive, answers. Here are a few examples:
These questions may provide some information on how the DIR performed but are ultimately inconclusive about the accuracy of the deformation.
The table below has some of the most common DIR QA methods used in clinics today. While you may use none, one, or a combination of them, let’s examine how each method answers the question of DIR accuracy.
|DIR QA Method||Does it Answer the Critical Question?||Why or Why Not?|
|Do Nothing||No||By placing absolute trust in the deformable image registration process, we fail to answer the critical question, thus elevating the risk of sending an inaccurate registration for planning and wasting time during later treatment planning stages.
|Deformed Image/Plan CT Fusion||No||Just because a deformed image looks like the planning CT, does not mean that local anatomy was correctly mapped between the two scans. While this method allows the user to quickly see if bone is mapped to soft tissue, it is extremely difficult—if not impossible—to determine whether soft tissue is mapped to the same soft tissue. Homogenous anatomy is equally, if not more important, than heterogenous anatomy.
The proposed method of looking for warping or stretching in the deformed image can also be erroneous. The amount of movement may have been necessary and correct. Its presence alone does not tell us whether the deformation is accurate or inaccurate.
Ultimately, this method fails to answer the critical question confidently and efficiently for all types of anatomy. It can, in fact, be misleading.
|Vector and Heat Maps||No||While vectors show the direction of movement and heat maps show the magnitude of movement, both fail to confirm if the voxels between two images correspond accurately at crucial points of interest for planning. Each end of the vector does not indicate the starting or ending point of a cluster of voxels. The magnitude of movement is not indicative of an inaccurate deformation, either. Sometimes a large amount of movement is necessary.
|Jacobian Determinant and Vector Curl||No||Jacobian Determinant values indicate how the volume of a voxel changes due to the transformation from the primary to secondary series in the registration. The Vector Curl measures the rotational forces applied to a voxel as it moves through the deformation field. Other than a Jacobian Determinant of 0, which is a non-physical movement and rare, these values can’t tell whether the deformation accurately linked a local cluster of voxels.
|Point-to-Point||Yes||Selecting points to verify that a cluster of voxels in one image corresponds to the same cluster of voxels in the other image, will spot errors regardless of the type of anatomy and the amount of movement. It is difficult, however, to identify precise locations when your eyes are moving back and forth between images. Unfortunately, it is a time-consuming and inefficient process.
From the table above, it may seem that point-to-point is the best choice for answering the critical question. Unfortunately, the point-to-point method is cumbersome and impractical. Clinicians don’t have the time to select numerous points on their secondary image and see where they correspond to their planning image.
Although the other QA methods have a certain level of convenience, they do not answer the most critical question. Even a combination of these methods still fails to produce accurate QA.
The most useful and accurate QA method would leverage point-to-point specificity with practical efficiency. When held to the critical question, it’s clear that a paradigm shift in DIR QA is needed to increase trust in the clinical use of deformable image registration.
If you'd like to learn more about MIM Software and how we answer the critical question in DIR QA, contact email@example.com.
Jeff Kuhn is a Product Manager at MIM Software. Jeff works closely with clinics across the country to improve their pre-treatment processes such as image registration, contouring, 4D motion management, and plan assessment.
You can visit them online at
The use of AI technology has transformed the field of Radiation Oncology treatment planning. This...