Non-Contrastive Learning for CT–CBCT Patient-Plan Mismatch Detection In Head and Neck IGRT
Abstract
Purpose
The incorrect patient placed on the radiotherapy couch for treatment is rare, but the consequences are potentially devastating. We developed an artificial intelligence assistance tool intended to support therapists at the treatment console to ensure the correct patient is receiving treatment.
Methods
We collected the simulation CT (simCT), cone beam CT (CBCT) and the associated clinical registration DICOM files from patients treated with head and neck radiotherapy at our institution between January 2016 and December 2022. We utilized a self-supervised approach inspired by SimSiam, a non-contrastive Siamese representation learning network. The two branches share an encoder network's weights to project paired inputs into the feature space. A predictor MLP is applied to one branch, and training maximizes the similarity between two views using negative cosine similarity loss. Clinically registered simCT and CBCT volumes served as the paired views, and we trained the network to learn robust representations of anatomic agreement between the paired views. The training, validation, and test sets included 4104, 954, and 970 registered volumes, respectively. 20 wrong-patient cases were created by manually registering the simCT and CBCT from different patients with similar anthropometric characteristics. At test time, negative cosine similarities from both passes were combined using linear discriminant analysis to form the anomaly detection score. We theorized wrong-patient registrations would generate embedding disagreement and feature misalignment, thus producing higher anomaly scores. Our method was benchmarked against traditional machine learning methods.
Results
Our anomaly detection method achieved an AUC of .986, while a simple linear discriminant using the mean-squared-error (MSE) between the volumes demonstrated an AUC of 0.883.
Conclusion
The robust performance of our algorithm shows the feasibility of a self-supervised approach to prevent treatment of the wrong patient, and we believe this tool can be utilized as an automated second check at the treatment console.