Head-and-neck IMRT presents a beam angle selection problem that is qualitatively different from most other treatment sites. In prostate IMRT, the PTV and the critical OARs — rectum, bladder, femoral heads — are geometrically separated in ways that allow standard coplanar beam arrangements (five to nine equally spaced beams around the axial plane) to achieve acceptable plans in most anatomical configurations. In head-and-neck cases, the geometric overlap between targets and critical structures is dense: the spinal cord runs posteriorly through the irradiation field, bilateral parotids sit immediately lateral to the high-dose volumes, the mandible and oral cavity are often partially within target regions, and bilateral neck node chains extend the PTV craniocaudally across a large anatomical range.
Beam angle selection in this context is not a minor planning detail. The choice of beam angles determines the geometric access to each target sub-volume, the extent to which posterior beams contribute to spinal cord dose, and the ability to avoid parotid tissue with oblique versus perpendicular beam entry angles.
The Standard Coplanar Arrangement and Its Limitations
Most IMRT head-and-neck protocols in Japan and internationally use coplanar beam arrangements — all beams in the axial plane at the isocenter — with gantry angles distributed around the 360-degree arc to avoid direct posterior beam entry on the spinal cord. A common nine-beam arrangement uses angles spaced at approximately 40-degree intervals, with a gap in the posterior sector to keep beams away from the cord. This arrangement is straightforward to plan, produces consistent results with experienced planners, and is the institutional standard at most departments.
The limitation of coplanar arrangements for bilateral neck IMRT is that all beam access is from the axial plane, which constrains the geometric options for differentiating dose to left versus right neck node levels and limits the geometric separation between parotid tissue and beam entry trajectories. When one parotid is partially overlapping the PTV due to node involvement, axial-plane beams cannot avoid that parotid without compromising coverage of the adjacent node level.
Non-coplanar beams — with couch rotations that shift beam entry above or below the axial plane — add degrees of freedom that can improve parotid avoidance and spinal cord sparing simultaneously. The cost is treatment delivery complexity: non-coplanar plans require couch rotations during treatment, which add setup time and couch positioning accuracy requirements. In departments with high patient throughput, the per-fraction time cost of non-coplanar beam delivery is a real operational constraint.
Automated Beam Angle Selection: Problem Formulation
Automated beam angle optimization (BAO) formulates beam angle selection as an optimization problem: find the set of angles that minimizes some combination of OAR dose and maximizes PTV coverage, subject to geometric and delivery constraints. The challenge is that beam angle space is combinatorially large — for a nine-beam plan chosen from 72 candidate angles at 5-degree resolution in the axial plane, the number of possible combinations exceeds 10 billion — and the objective function is expensive to evaluate because each candidate angle set requires a full inverse planning run to assess dose quality.
Practical BAO algorithms address this combinatorial problem through approximation strategies: column generation methods that iteratively add beams to a growing plan; sequential beam selection that adds beams one at a time based on marginal dosimetric contribution; or machine learning approaches that predict beam quality from geometric features without running the full optimizer for each candidate set. Each approach makes different accuracy-speed tradeoffs; none produces provably optimal beam configurations, but all produce configurations that are competitive with or superior to standard fixed-angle arrangements when evaluated on clinical test cases.
The most clinically relevant improvement from BAO in head-and-neck cases is not the identification of exotic angle combinations that human planners would never consider. It is the systematic exploration of angle perturbations around the standard arrangement — evaluating whether shifting the anterior oblique beams by 10–15 degrees, or adding a single non-coplanar beam from a superior direction to improve mandible avoidance, produces a meaningfully better plan. These are adjustments that experienced planners make intuitively on difficult cases; BAO makes the same exploration systematic and reproducible across planners.
Parotid Geometry and the Ipsilateral Constraint Problem
Consider a specific scenario: oropharyngeal carcinoma with T2N2b staging, left-sided primary with bilateral neck node involvement extending to level IV. The PTV encompasses bilateral levels II-IV and the primary with a standard 5 mm CTV-to-PTV expansion. The left parotid is partially within the PTV at level II due to the nodal volume. The right parotid is geographically separated but receives dose from beams traversing the right neck.
In this geometry, the left parotid cannot be fully spared — part of it is inside the prescription-dose volume by definition. The clinically relevant objective is to spare as much of the remaining (non-overlapping) left parotid as possible while fully covering the node. For the right parotid, the objective is mean dose minimization. These are distinct optimization problems with distinct geometric solutions.
BAO can improve right parotid sparing by identifying beam entry trajectories that access the right neck node levels without traversing the right parotid — typically through anterior and anterior-oblique angles that provide node coverage while beam exit routes avoid the lateral parotid. For the left parotid, the BAO improvement is more limited because the geometric constraint is the PTV overlap, not the beam trajectory. An AI planning system that understands this distinction — treating the right parotid as a standard BAO target and the left parotid as a contiguous-PTV case requiring a different constraint formulation — produces more clinically relevant optimization than one that applies the same treatment to both structures.
Volumetric Modulated Arc Therapy vs. Step-and-Shoot IMRT
Most head-and-neck IMRT in Japan is now delivered as VMAT (volumetric modulated arc therapy) rather than step-and-shoot IMRT. VMAT's continuous arc delivery changes the beam angle optimization problem: instead of selecting discrete gantry angles, the optimizer controls MLC positions, dose rate, and gantry speed as continuous functions around the arc. The arc start angle, the arc span, and the number of arcs (single vs. dual arc) are the configurable parameters that correspond to beam angle selection in static IMRT.
For bilateral head-and-neck cases, dual-arc VMAT (two full or partial arcs covering the 360-degree range) typically provides better OAR sparing than single-arc delivery because the two arcs can be assigned different gantry angle ranges that preferentially access different sub-volumes. The posterior arc sector — angles near 180 degrees — requires careful MLC speed management to limit spinal cord contribution; AI-based VMAT optimization that incorporates arc angle and sector weighting as learnable parameters rather than fixed protocol values can improve outcomes on anatomically complex cases.
We are not saying that automated beam angle optimization eliminates the need for physicist expertise in VMAT planning — the arc angle and sector decisions interact with MLC optimization and delivery constraints in ways that require physicist oversight to evaluate. The physicist reviewing a BAO-generated VMAT plan for a complex head-and-neck case needs to understand why specific arc sectors were weighted the way they were, not just whether the DVH metrics are satisfied.
What Remains Genuinely Difficult for Automation
Beam angle optimization for head-and-neck IMRT automates well the systematic search through angle space. What it cannot automate is the clinical judgment about which anatomical features of this specific patient's imaging should drive the angle selection: is the imaging-apparent node volume reliable, or is there contour uncertainty that affects the PTV extent near the parotid? Is the patient's typical treatment position stable enough for non-coplanar delivery, given their reported neck mobility after previous surgery? Does the treating oncologist have a strong preference for a delivery arrangement that is familiar to the therapist team, overriding a marginally superior BAO result?
These are not edge cases that better algorithms will eventually solve. They are the structural features of clinical decision-making that require human involvement at every planning session. The appropriate role of BAO in head-and-neck IMRT is to reduce the time physicists spend on combinatorial angle search and to surface better configurations than fixed-arrangement protocols, while the physicist applies clinical judgment to evaluate those configurations in the context of each patient's situation.