Purpose: The design of a high-quality treatment plan is often very time-consuming as it involves tuning tradeoffs. Retrieving similar patients from a database of previously treated patients can facilitate treatment planning, but it is difficult to achieve due to the complicated relationship between RT structure and dose distribution. Here, we utilize Relational Autoencoder (RAE) to develop a novel relationship descriptor considering both patient geometry and the relationship of dosimetric properties of different patients for retrieving patients similar to the query patient.
Methods: A RAE model is employed to obtain a low-dimensional feature of 3D RT Structures for retrieving similar patients, meanwhile maintain the relationship of dosimetric similarity among patients in feature space. The geometric consistence is achieved by minimizing the Dice-coefficient loss for the reconstructed contour masks of a single sample, and the dosimetric relationship coherence is realized by minimizing the mean square error (MSE) of the Pearson linear correlation coefficient (PLCC) of Dose-Volume Histograms (DVH) and encoded features for each batch of training samples. The feature is then used as the relationship descriptor for retrieving similar patients to a query patient.
Results: The RAE is trained with DVHs and RT Structures of 76 prostate patients and tested on 8 patients. The encoder of RAE compresses the contour mask images of 3D RT Structures into a descriptor of 276 numbers, and the geometry of the RT Structures is well reconstructed through the decoder. The proposed relationship descriptor outperforms OVH based shape descriptor in retrieving patients with radiotherapy treatment plans similar to the query patient.
Conclusion: We have utilized RAE to obtain a novel relationship descriptor for retrieving similar patients in radiotherapy treatment planning. The results demonstrate that the task-specific RAE-based relationship descriptor outperforms traditional shape descriptors.