Radiation Therapy Planning: How AI Improves Target Volume Delineation

The integration of PET/CT into radiation therapy planning has changed how radiation oncologists define target volumes for several cancer types. FDG PET identifies the metabolically active gross tumor volume (GTV) with functional specificity that CT alone cannot provide - distinguishing viable tumor from post-obstructive atelectasis in lung cancer, active disease from treatment-related changes in lymphoma, and FDG-avid disease extension beyond anatomically apparent boundaries in head and neck cancer. AI-assisted target volume delineation is now addressing the most consistent problem with PET-guided radiotherapy: that manually contoured PET-based GTVs show inter-observer variability that would be considered unacceptable if it occurred with CT contouring.

The Inter-Observer Variability Problem in PET-Based Contouring

CT-based GTV contouring in well-defined solid tumors (lung peripherally located, clearly demarcated head and neck primary) shows inter-observer variability of 10-20% in volume between experienced radiation oncologists. This is considered the baseline challenge that contouring guidelines and atlases attempt to reduce. FDG PET-based GTV contouring in the same tumors shows inter-observer variability of 25-40% - substantially higher, despite the addition of metabolic information that theoretically should reduce ambiguity.

The paradox has a clear explanation: PET images lack clear boundaries. A CT-visible mass has a measurable edge based on Hounsfield unit change. A PET GTV boundary is defined by a threshold of metabolic activity - but no consensus exists on what that threshold should be, or whether it should be a fixed threshold (e.g., SUV 2.5), a percentage of maximum uptake (40% of SUVmax is a common approach), a gradient-based algorithm, or a visually determined contour. Different oncologists using different threshold definitions on the same PET study produce different GTVs, and those differences translate to different doses delivered to different volumes of tissue surrounding the tumor.

This variability is not merely academic. A patient whose PET-based GTV is contoured smaller than another physician would contour it receives lower dose to the peripheral tumor regions. Whether that peripheral region contains viable tumor or not depends on the tracer uptake characteristics of the specific tumor type - but the average case receiving under-dosed peripheral disease from threshold selection variability is a real clinical concern, not a hypothetical one.

Automated Segmentation Approaches: Threshold vs. Deep Learning

The simplest automated PET GTV segmentation approach is threshold-based: identify all voxels with SUV above a specified threshold, connected to the primary lesion, as the GTV. The threshold can be fixed, adaptive (based on lesion SUVmax), or background-relative (based on the SUV ratio of tumor to adjacent tissue). Threshold-based segmentation is reproducible - the same threshold produces the same contour on the same image - but the choice of threshold remains arbitrary and tumor-type dependent.

Published data comparing threshold-based segmentation against manual contouring with pathological ground truth (resected specimens with histology correlation) shows that no single threshold value reliably reproduces the true disease volume across tumor types. A 40% SUVmax threshold that performs acceptably for FDG-avid lymphoma GTVs systematically underestimates volume in moderately FDG-avid lung adenocarcinoma with a heterogeneous uptake pattern.

Deep learning segmentation models trained on paired PET/CT images with expert-contoured GTVs learn the relationship between multi-modal image features and tumor boundary location - incorporating spatial context, gradient information, and uptake heterogeneity patterns that fixed-threshold approaches cannot capture. Models trained specifically on head and neck cancer PET/CT data with consensus contours from multiple radiation oncologists achieve Dice similarity coefficients of 0.75-0.82 for primary GTV and 0.68-0.74 for nodal GTV - acceptable performance for clinical use with physician review.

Dose Painting: The Next Frontier

Standard radiation therapy delivers uniform dose to the defined target volume. Dose painting is an approach that modulates dose within the target based on imaging-derived biological information - delivering higher dose to the highest-SUV subvolumes within a tumor on the premise that these represent the most radioresistant disease (hypoxic, rapidly proliferating, or otherwise challenging tissue).

PET-based dose painting requires automated subvolume delineation within the GTV - identifying the high-risk biological target volume (BTV) within the broader GTV and specifying it as the dose-boost region in the treatment plan. Manual subvolume delineation is extremely difficult because PET image noise at the voxel level makes reliable identification of high-uptake regions within the tumor challenging. AI-based subvolume segmentation, using algorithms that incorporate spatial smoothing and noise estimation, provides more reproducible BTV definitions than manual approaches.

Clinical evidence that dose painting to PET-defined BTVs improves tumor control compared to uniform dose delivery is still accumulating. The RTOG 1106 trial for lung cancer and the PET-Boost trial in the Netherlands investigated dose escalation to PET-avid subvolumes - results have been mixed, with some benefit seen in specific subgroups and none in others. The concept is biologically rational; the challenge is that PET at standard resolution may not provide the spatial precision needed to reliably target subcentimeter high-risk subvolumes with the dose conformality achievable by modern IMRT or SBRT techniques.

Head and Neck Cancer: The Strongest Clinical Use Case

Head and neck squamous cell carcinoma represents the cancer site with the strongest evidence for PET-based target volume delineation affecting treatment quality. The primary tumor volume and involved lymph nodes in head and neck cancer are surrounded by adjacent critical structures - parotid glands, spinal cord, mandible, pharyngeal constrictors - where radiation dose must be carefully managed to avoid late toxicities including xerostomia, dysphagia, and osteoradionecrosis.

FDG PET identifies the active tumor extent with sensitivity that detects disease beyond the CT-visible boundary in 15-20% of cases, allowing margin reduction in the CT-based GTV while including biologically active disease that CT misses. Prospective randomized trial data (the PET-PLAN trial) showed that PET-guided target volume reduction in head and neck cancer achieved equivalent tumor control with significantly lower doses to parotid glands - a direct late-toxicity benefit from imaging-guided contouring.

AI-automated PET GTV delineation in head and neck cancer has a direct practical impact on this benefit: it makes PET-guided planning faster, more reproducible, and feasible at centers that currently do not incorporate PET into every head and neck treatment plan due to contouring time constraints.

Limitations of AI Delineation That Are Rarely Discussed

AI delineation tools for radiotherapy are increasingly marketed on their Dice similarity coefficient performance against reference contours - a metric that measures geometric overlap but does not directly measure clinical relevance. A contour that is geometrically accurate on average can still be clinically wrong in specific cases: overextending into an adjacent critical structure, missing a portion of the tumor at the interface with post-obstructive disease, or including physiological FDG uptake misidentified as tumor.

The physician review step after AI delineation is therefore not optional and should not be reduced to a checkbox. The cases where AI delineation fails are precisely the cases with unusual tumor morphology, adjacent structure complications, or tracer uptake patterns not well represented in the training data - which are also the cases most likely to be high-stakes. A standardized review workflow that focuses physician attention on regions of uncertainty - boundary areas, adjacency to critical structures, regions of heterogeneous uptake - is more appropriate than wholesale acceptance of AI contours.