Bagher-Ebadian H, Nagaraja TN, Cabral G, Farmer K, Valadie OG, Acharya P, Movsas B, Brown SL, Chetty IJ, and Ewing JR. An Unsupervised Autoencoder Developed from Dynamic Contrast-Enhanced (DCE)-MRI Datasets for Classification of Acute Tumor Response in an Animal Model. Int J Radiat Oncol Biol Phys 2022; 114(3):S34.
Int J Radiat Oncol Biol Phys
Purpose/Objective(s): Recent studies have shown that vascular parameters of brain tumors derived from DCE-MRI may act as potential biomarkers for radiation-induced acute effects. However, accurate characterization of the spatial regions affected by radiation therapy (RT) remains challenging. Here, we introduce an unsupervised adaptive model for classification and ranking of the RT-affected regions in an animal model of cerebral U-251n tumors.
Materials/Methods: Twenty-three immune-compromised-RNU rats were implanted with human U251n cancer cells to form an orthotopic glioma (IACUC #1509). For each rat, 28 days after implantation, two DCE-MRI studies (Dual Gradient Echo, DGE, FOV: 32 × 32 mm2, TR/(TE1-TE2) = 24 ms/(2 ms-4 ms), flip angle = 18°, 400 acquisitions, 1.55 sec interval with Magnevist contrast agent, CA injection at ∼ 24 sec) were performed 24h apart using a 7T MRI scanner. A single 20 Gy stereotactic radiation exposure was performed before the second MRI, which was acquired 1-6.5 hrs after RT. DCE-MRI analysis was done using a model selection technique to distinguish three different brain regions as follows: Normal vasculature (Model 1: No leakage, only plasma volume, vp, is estimated), leaky tumor tissues with no back-flux to the vasculature (Model 2: vp and forward volumetric transfer constant, Ktrans, are estimated), and leaky tumor tissues with back-flux (Model 3: vp, Ktrans, and interstitial volume fraction, ve, are estimated). Normalized time traces of DCE-MRI information (24 pre, and 24 post-RT for each rat, total of 64108 training datasets) of tumors and their soft surrounding normal tissues were extracted from the 3 different model regions. To eliminate high-dimensional data similarity, an unsupervised autoencoder (AE) was trained to map out the model-derived data into a feature space (latent variables, N=10). For each model, the pre and post RT latent variables were compared (by appropriate tests of significance: ANOVA/Welch, CI=95%) to reveal RT-discriminant features. Pearson correlation coefficients were used to compare the decoded data to rank the effect of RT on different models.
Results: The time trace of DCE-MRI information of rat brain in normal (Model 1, non-leaky) and highly permeable (Model 3) regions are less impacted by RT (Higher correlation between pre and post RT: r= 0.8518, p<0.0001 and r= 0.9040, p<0.0001 for Model 1 and Model 3, respectively) compared to the peritumoral regions pertaining to Model 2 (r= 0.8077, p<0.0001).
Conclusion: This pilot study suggests that among different brain regions, peritumoral zones (infiltrative tumor borders with enhanced rim) are highly affected by RT. Spatial assessment of RT-affected brain regions can play a key role in optimization of treatment planning in cancer patients, but presents a challenging task in conventional DCE-MRI. This study represents an important step toward classification and ranking the RT-affected brain spatial regions according to their vascular response following hypofractionated RT.