2052 - Multiple-Channel Hybrid Neural Network to Achieve Quantitative Bioluminescence Tomography for Cancer Research
Presenter(s)
B. Deng1, Z. Tong1, J. Kim2, H. Dehghani3, and K. K. H. Wang1; 1Biomedical Imaging and Radiation Technology Laboratory (BIRTLab), Department of Radiation Oncology, UT Southwestern Medical Center, Dallas, TX, 2Department of Internal Medicine, UT Southwestern Medical Center, Dallas, TX, 3School of Computer Science, University of Birmingham, Birmingham, United Kingdom
Purpose/Objective(s): Bioluminescence tomography (BLT) enhances commonly used 2D BL imaging by reconstructing 3D distributions of BL activity, allowing in vivo tumor localization. However, conventional model-based BLT faces ill-posedness, model and data uncertainties, limiting its ability to provide quantitative and robust information, such as tumor volume. To address these, we introduce the Multiple-channel Hybrid Neural Network (McHyNN), integrating light propagation model, multimodal imaging, and cross-animal data. With McHyNN, by incorporating the contours of MRI-detectable large tumors, we expect that we can significantly improve BLT accuracy in localizing small tumor that could be missed by MRI. This advancement of McHyNN potentially positions BLT as a promising quantitative imaging to address challenging cancer research problems, such as investigating spontaneous metastasis in animal models.
Materials/Methods: McHyNN is a feed-forward neural network (NN) grounded in light diffusion model. It incorporates multi-data channels, including multi-animal and multimodal data. For a given animal set, the data includes BL images emitted from internal tumors and corresponding MRI scan, if tumor is MRI-detectable. The NN outputs n+1 tumor distributions, including the n existing animal sets, and the new subject whose tumor location and volume are to be determined. McHyNN leverages mutual information from light propagation across multiple animals and multimodal images, and the NN is optimized accordingly by loss function L = Ld + Lim. Here, Ld is the difference between calculated and measured surface BL data using a spectral-derivative approach to address data uncertainty, while Lim accounts for MRI constraints. We will first present numerical results based on simulated brain metastases(BMs) in mice with experimentally determined noise, followed by showing the capability of BLT to reconstruct in vivo BMs either using intracranial injection or SCLC spontaneous BMs model.
Results: McHyNN accurately quantifies 3 small BMs (volume 0.9-1.9 mm3) within simulated mouse brain at localization uncertainty<0.5mm and volume errors<0.8 mm3, even when the tumors are separated by only 2–2.7 mm, while the conventional BLT fails to distinguish these tumors. Moreover, we simulated single BM ranging from 0.4-17.2 mm3, including 10 tumors<1.8mm3 that are often challenging for MRI to detect. McHyNN successfully localized all cases <0.5 mm location accuracy and faithfully reconstructed tumor volumes, showing a linear correlation with ground truth (R2 =0.99). Notably, MRI data was only available for larger tumors, yet McHyNN utilized this information to recover smaller tumors—an unprecedented achievement in optical tomography. The in vivo BMs study is ongoing.
Conclusion: By incorporating the McHyNN, BLT is promising to achieve quantitative imaging, enabling precise tumor volume quantification and detection of submillimeter malignancies to enhance investigators’ capabilities in cancer research.