Advancing Radiotherapy with AI-Driven Outlier Detection

Abstract

In radiation oncology, there is a strong and growing research focus on deep learning methods for automated treatement planning. These models rely on large amounts of high quality clinical training data in order to generate reliable dose predictions and treatment plans. However, radiotherapy datasets typically consist of heterogeneous Digital Imaging and Communications in Medicine (DICOM) modalities, including computed tomography(CT) images, structure sets, and treatment plans, which may contain anatomical anomalies,technical artifacts, contouring in consistencies, or protocol deviations. Since they arefor individual patient treatment, not for depp learning tasks. The aim of this thesis was therefore to develop and evaluate a multi-modal anomaly detection pipeline that actsas an automated “Quality Gate” for radiotherapy datasets before their use in Artificial Intelligence (AI)-driven treatment planning workflows.To address the different characteristics of the radiotherapy modalities, multiple detection strategies were implemented. CT images were analyzed with a continuous Variational Autoencoder (VAE) baseline and a discrete Vector Quantized Variational Autoencoder (VQ-VAE). Structure sets were evaluated using both a voxel-based StructureVAE and an geometric Principal Component Analysis (PCA) detector. Treatment plans were analyzed with a deterministic rule-based detector in order to flag protocol deviations and hardware-related out-of-distribution parameters.The developed pipeline was evaluated on a clinical prostate cancer cohort consisting of190 patients. The data was split into a training and validation split using only normal patients, while the test split consisted of 12 normal patients and 27 labeled outliers.For the CT-modality the VQ-VAE outperformed the baseline and for structuresets,PCA out performed the Structure Variational Autoencoder (StructureVAE). For the plan detection the detector identified relevant protocol and hardware deviations that were not directly visible in the other modalities.The results displays that ure deep learning approaches are not sufficient for robust radiotherapy data curation. Instead, hybrid systems that combine deep learning, statistical modelling, and domain rules are better suited to detect anomalies. The curated dataset improved downstream model performance, with an 18% improvement in dose-prediction training performance and a 9.5% reduction in plan-generator validation loss. These results indicated that automated outlier detection can improve data quality and improve AI-assisted radiotherapy planning.