Combining autoencoder with clustering analysis for anomaly detection in radiotherapy plans.

Abstract

To develop an unsupervised anomaly detection method to identify suspicious error-prone treatment plans in radiotherapy. A total of 577 treatment plans of breast cancer patients were used in this study. They were labeled as either normal or abnormal plans by experienced clinicians. Multiple features of each plan were extracted and selected by the learning algorithms. The training set consisted of feature samples from 400 normal plans and the testing set consisted of feature samples from 158 normal plans and 19 abnormal plans. Using the k-means clustering algorithm in the training stage, 4 normal plan clusters were formed. The distance between the samples in the testing set and the cluster centers were then determined. To evaluate the effect of dimensionality reduction (DR) on detection accuracy, principal component analysis (PCA) and autoencoder (AE) methods were compared. The sensitivity of the anomaly detection model based on PCA and AE methods were 84.2% (16/19) and 94.7% (18/19), respectively. The specificity of the anomaly detection model based on PCA and AE methods were 64.6% (102/158) and 69.0% (109/158), respectively. The areas under the receiver operating characteristic (ROC) curve (AUCs) based on PCA and AE methods were 0.81 and 0.90, respectively. The unsupervised learning method was effective for detecting anomalies from the feature samples. Accuracy could be improved with the introduction of AE-based DR technique. The combination of AE and k-means clustering methods provides an automated way to identify abnormal plans among clinical treatment plans in radiotherapy.