Motion during treatment can be tracked with cameras or IR sensors or subcutaneous probes but that doesn't tell you about internal organs moving. The topic of deformable registration, where you find a non-rigid mapping between initial imaging conditions and the current ones, is still a topic of active research. Adaptive planning, where you actively change the treatment plan every N sessions based on the most up to date information, is also actively researched / implemented in some good research centers.
For treatment planning you just use a standard Cartesian grid, or a "beam's eye view" coordinate system that's aligned with the radiation beam axis as it rotates around the patient.
Not to be glib but on behalf of the thousands of people going into a radiotherapy clinic today for treatment, thanks for working to improve these techniques.
People in my field are fairly fortunate as there is a career track as a clinical medical physicist that is highly paid and pretty low stress, so most people end up going there. The work consists of maintaining and calibrating the radiation therapy machines, along with implementing new technologies in the clinic, and fixing problems that don't fall within the job description of the radiation therapists. Like what to do when a radioactive seed falls on the floor instead of going inside the patient where it's supposed to go. There's also a separate track as an imaging physicist where you maintain and QA the diagnostic imaging machines.
I'm personally doing a postdoc at the junction between optimisation, machine learning and radiation therapy. Just starting out though. Basically just extending my PhD work to automate the treatment planning process and remove the variability in treatment plan quality due to the level of experience of the people making the plans.