Session Item

Daily MRIs acquired during MR-guided radiotherapy (MRgRT) provide a unique source of data for AI investigation. Analysis of morphological changes within the tumor enables tumor behavioral prediction that can lead to improvements in treatment efficacy. We hypothesized that changes in gross tumor volumes (GTVs) analyzed through daily MRI acquisition through a novel deep neural network (DNN) architecture would provide the best predictive power for disease progression compared to DR (Delta Radiomics) with external validation.

We analyzed 65 patients, with 390 MRI scans, treated on a 0.35T MR linac across two institutions from 2019 to 2021. Internal datasets (IDS) composed of a multi-disease model from adrenal and lung tumor patients (N=47) from a single institution, and external dataset (EDS) of adrenal cancer for validation from an outside institution (N=18). Fixed-bin number (FBN) intensity discretization was performed on the GTV volumes using 64 bins. A DNN was developed using 3D residual network blocks followed by blocks of transformers for imaging and temporal feature extraction. DNN training strategies included training-from-scratch, sequential transfer learning, and a novel multi-task (parallel) learning (MTL) approach. All model optimization, selection, and training were performed exclusively on the IDS. Evaluation of predictions was assessed using area under the ROC Curve (AUC) and with 1,000 iterations of the Bootstrap .632+ method. We then compared the DNN model with DR strategies. Six DR classification models with 73 texture features were extracted and trained via feature ratios between the first and last fractions as per previous works by our group (Tomaszewski et. al. 2021). The EDS was used only for independent testing of the final models via level 3 criteria of the transparent reporting of a multivariable prediction model for individual prognosis or diagnosis (TRIPOD).

Each patient had 6 scans, 1 at sim and 1 at each fraction. GTV was defined on sim and co-registered to scans. Median RT was 50 Gy in 5 fractions for all tumors and the median follow-up of 10 months (adrenal) or 17 months (lung) of the IDS training set. Tumor progression was seen in 16 IDS adrenal, 11 IDS lung, and 9 EDS adrenal patients. DNN and DR performances are summarized in table 1. All hybrid DNN models demonstrate superior predictive power over DR. TRIPOD demonstrated that the MTL DNN model had the strongest external predictive potential for predicting disease progression with an AUC of 0.876 on external validation.

This proof-of-concept model demonstrates that a multi-task (parallel) learning DNN has superior power to predict for tumor progression during MRgRT compared to traditional DR. This hybrid multi-task DNN is validated on an external dataset with a high AUC. This opens the door for real-time physiologic-response adaptation strategy via MRgRT in clinical trial development.