(E-180) Machine Learning of Vascular Pharmacokinetic Parameters and Permeability Changes in Brain Tumors

Introduction:: Glioblastoma multiforme (GBM) and brain metastasis (BM) are the most common and aggressive intracranial malignances in adults. Vascular permeability and perfusion has been shown to be markedly heterogeneous in brain tumors due to many of these lesions retaining an intact blood-tumor barrier (BTB). Quantitative MRI methods such as dynamic contrast enhanced (DCE) MRI are becoming increasingly common in BTB permeability assessment as well as for assessment of permeability responses to treatment. However, conventional pharmacokinetic (PK) modeling in DCE MRI is complex and time consuming for dynamic MR scans with thousands of pixels per image. In the present study, we have developed deep learning approaches using convolutional neural networks (CNN) [1] and transfer learning (TL) for the generation of PK parameter maps in small animal models of GBM and BM. Furthermore, the utility of the CNN to predict permeability changes was investigated in radiotherapy treated BM lesions. We are currently investigating the translational potential of the CNN to predict vascular permeability and treatment responses in clinical brain tumors.


Materials and Methods:: Orthotopic human GBM tumors in nude mice (n=6) were established through intracranial injections of U87 cells. Intracardiac injections of MDA-MB-231-BR BM cells under ultrasound guidance was used to develop a small animal model of breast cancer BM in nude mice (n=16). MR imaging was performed on a 7T Bruker Biospin small animal scanner to monitor tumor development. Once tumor development was established, anatomic imaging and DCE MRI was conducted. A subset of BM mice (n=10) were randomly assigned to either whole brain radiotherapy (WBRT) or sham treatment groups. WBRT was conducted in three daily fractionations of 4 Gy using an X-RAD 320 orthovoltage irradiator. MR imaging was then acquired following treatment. A 24-layer CNN was designed and trained to learn the relationship between GBM DCE images and target permeability PK maps. A deep learning approach using TL was then employed through re-training of a GBM-trained CNN with BM MR imaging datasets. K-fold cross validation was employed in both CNN and TL studies to allow all animals to serve as both training and testing data.


Results, Conclusions, and Discussions:: As shown in the representative cases in Figure 1, high-resolution T2-w imaging (T2-WI) revealed a large solitary hyperintense lesion in the GBM model with high enhancement on T1-CE images indicating BTB leakage. Unlike the glioma model, T2-WI revealed hyperintense multifocal metastases in the BM model including both contrast enhancing and non-enhancing lesions on T1-CE images. Clearly, the CNN models could accurately assess permeability in both GBM and BM and presented with excellent spatial correlation to target PK maps of permeability parameter Ktrans extracted from the Extended Tofts Model (ETM). Furthermore, the CNN detected significantly increased Ktrans in WBRT treated BM tumors (p< 0.01), which was in high agreement with the target ETM PK maps (p< 0.01).

In conclusion, the proposed CNN and TL approaches can serve as efficient and accurate tools for characterizing vascular permeability and treatment responses in small animal brain tumor models. Further investigations of the utility of the machine learning approach to predict vascular permeability and radiotherapy treatment responses in clinical brain tumors are underway.



Acknowledgements (Optional): : This research was supported in part by NIH/NCI 1R01CA194578 and Wake Forest CCC P30 CA01219740.


References (Optional): : [1] Arledge CA et al. Front Biosci (Landmark Ed). 2022. 27(3):99.