Purpose: Clinical target volumes (CTV) are contoured for postoperative prostate radiotherapy to treat microscopic disease with the aid of consensus guidelines. Adjustments are made in accordance with a physician’s preference/clinical judgement, resulting in large observer-variations (segmentation-styles) that are distinguishable. Although autosegmentation models can be adapted to specific physician styles, this would require sufficient data from each physician for their styles to be learnable. To address this limited sample size issue, we propose a deep learning (DL)-based few-shot meta-learning architecture for CTV style adaptation.
Methods: We used existing prior information for a specific CTV contouring style in the form of previously segmented patients as support dataset. To train the meta-learning architecture, we simulated 6 different CTV styles (4-training, 2-testing) by varying the inclusion of bladder/rectum/seminal-vesicles. Initial CTV segmentations were performed via an in-house DL model. Then, our adaptive learning architecture was tasked to learn the changes required for new CTV styles. Three-hundred and seventy-three patients treated at a single institution were used. The network consisted of four parts: a shared encoder, style block, averaging layer, and decoder. The encoder extracted features from prior patients and the new patient, the style block evaluated the difference between the DL model CTV and the provided support style, and the average layer averaged features across all prior patients. These were concatenated with the new patient features and decoded to obtain the adapted CTV.
Results: The performance of the in-house CTV segmentation model on the two unexposed styles for 50 independent test patients were 71.9% and 63.0% dice score coefficient (DSC). With just three prior patients as input to the few-shot network, the mean DSC across test patients’ increased to 77.8% and 73.0% for each style, respectively.
Conclusion: The trained few-shot meta-learner can effectively adapt to new CTV styles.
Funding Support, Disclosures, and Conflict of Interest: We would like to thank Varian Medical Systems, Inc. for providing funding support.
Computer Vision, Segmentation, 3D