A general model for head and neck auto-segmentation with patient pre-treatment imaging during adaptive radiation therapy.

Background: During head and neck (HN) radiation therapy, patients may undergo anatomical changes due to tumor shrinkage or weight loss. For these patients, adaptive radiation therapy (ART) is required to correct treatment plans and to ensure that the prescribed radiation dose is delivered to the tumor while minimizing dose to the surrounding organs-at-risk (OARs). Patient pre-treatment images and segmentation labels are always available during ART and may be incorporated into deep learning (DL) auto-segmentation models to improve performance on mid-treatment images.

Purpose: Existing DL methods typically incorporate pre-treatment data during training. In this work, we investigated whether including pre-treatment data at inference time would affect model performance, as inference-time inclusion would eliminate the requirement for costly model retraining for new patient cohorts.

Methods: We developed a general adaptive model (GAM) that included pre-treatment data at inference time through additional input channels. We compared the GAM with a patient-specific model (PSM), which included pre-treatment data during training, a reference model (RM), which did not include pre-treatment data, and a rigid image registration (RIR) method. Models were developed using a large dataset of pre- and mid-treatment computed tomography images and segmentation labels (primary gross tumor volume [GTVp] and 16 OARs) for 110 patients who underwent ART for HN cancer.

Results: The GAM showed improved performance over the PSM and RM for several structures, with the largest differences in dice similarity coefficient for difficult-to-segment structures: the GTVp (RM: 0.17, PSM: 0.36, GAM: 0.61, RR: 0.65) and left/right brachial plexus (RM: 0.38/0.35, PSM: 0.43/0.43, GAM: 0.49/0.49, RR: 0.36/0.38). The GAM attained similar performance to RR for all structures except the brainstem (GAM: 0.82, RR: 0.74), mandible (GAM: 0.88, RR: 0.68), and spinal cord (GAM: 0.76, RR: 0.51), for which the GAM performed higher.

Conclusion: The inclusion of patient pre-treatment images and segmentation labels can improve auto-segmentation performance during HN ART, in particular for structures with high variability or low contrast. Including pre-treatment data at DL model inference time (GAM) may give improvements over standard DL models for the GTVp and several OARs, while eliminating the need for costly model retraining with new patient cohorts. However, rigid registration provides similar performance to adaptive DL models for the GTVp and most OARs.