FARE2018 WINNERS Sorted By Institute NIH Clinical Center Adam Harrison Postdoctoral Fellow Radiology/Imaging/PET and Neuroimaging Progressive and Multi-Path Holistically Nested Neural Networks for Pathological Lung Segmentation from CT Images Pathological lung segmentation (PLS) is an important, yet challenging, medical imaging application. Pulmonary diseases are a major source of death and hospitalization worldwide, for which computed- tomography (CT) is the leading analysis modality. As such, there is great impetus to develop CT tools for automated disease detection and diagnosis. Yet, such tools, in turn, often rely on reliable PLS, i.e., an accurate delineation of pulmonary regions. As such, methodological simplicity and generality are key factors in a PLS tool's usability. Any tool must also handle the wide variability in appearance and shape of pathological lungs. State-of-the-art methods lack either simplicity or robustness. Along those lines, we present a bottom-up deep-learning based PLS method that is expressive enough to handle variations in appearance, while remaining unaffected by any variations in shape. We build off a deeply-supervised and fully-convolutional network (FCN) architecture called holistically nested networks (HNNs). To address the well-known coarsening problem of FCNs, we enhance the architecture by a simple, yet effective, progressive multi-path scheme, which more reliably merges outputs from different network stages and resolutions. Unlike other solutions to this problem, our multi-path scheme requires no extra parameters, and even has less parameters than standard HNNs, ensuring our method remains straightforward and usable. This results in a deep model able to produce finer detailed masks, which we call progressive holistically-nested networks (P-HNNs). Depending on the number of slices, our tool only takes 10-30s to process an entire CT volume. Using extensive cross-validation, our method is tested on multi-institutional datasets comprising 929 CT scans (848 publicly available) of pathological lungs, which include cases of interstitial lung diseases, chronic obstructive pulmonary disease, and various infections. This is the largest and most rigorous analysis to date of a PLS tool's performance. We report mean Dice scores of 0.985, significantly (p<0.001) outperforming standard HNNs and providing high-quality masks in instances where HNNs completely fail. As well, we also compare against a prior state-of-the- art, but non-deep-learning, PLS tool, demonstrating significant qualitative and quantitative (p<0.001) improvements. These results demonstrate the potential of our tool to contribute toward large-scale and high-throughput analysis of medical scans. __________________________________________________________________________________ NIH Clinical Center Andrew Mikhail Research Fellow Radiology/Imaging/PET and Neuroimaging Drug dose mapping using imageable, drug-eluting embolic beads for transarterial chemoembolization in a preclinical rabbit tumor model An estimated 40 000 new cases of primary liver cancer are expected in the United States in 2017. Regrettably, only 10-20% of patients with hepatocellular carcinoma (HCC), the most common form of primary liver cancer, are candidates for surgery. Transarterial chemoembolization (TACE) is a minimally- invasive treatment that consists of selective catheterization of tumor-feeding arteries under fluoroscopic image-guidance and delivery of embolic materials that cut off blood supply to the tumor. Drug-eluting embolic beads (DEBs) containing the chemotherapeutic agent doxorubicin (DOX) have been developed that serve both as an embolic agent and drug delivery vehicle for sustained tumor-localized delivery of chemotherapy. Recent development of drug-eluting radiopaque beads (DEROB) that are visible by fluoroscopy and CT may allow for optimization of therapy based on imaging feedback regarding DEROB distribution and density. Moreover, the image-ability of DEROB raises an intriguing potential for drug dosimetry, whereby relative levels of attenuation (image contrast) on CT may act as a surrogate for drug dose distribution during TACE facilitating greater customization of therapy. The purpose of this study was to determine the correlation between DEROB x-ray attenuation on CT and DOX concentrations in the liver and to estimate drug dose following TACE in a preclinical tumor model. Rabbits bearing VX2 liver tumors underwent TACE with DEROB following which the livers were resected, frozen and imaged with a 16-slice multidetector CT (MDCT). For tissue sectioning, the frozen livers were inserted into liver-specific 3D printed molds containing cutting slots for precise radiologic-pathologic correlation of tissue sections and imaging. A linear correlation was found between DEROB attenuation in the liver determined by MDCT image segmentation, and the concentration of DOX measured by liquid chromatography (r2 = 0.971). Regression analysis of DOX predicted on CT vs. DOX measured in a subsequently treated liver demonstrated accurate and precise drug dose estimation (slope = 1.06, intercept = 2x10-4 mg DOX, R2 = 0.93). This relationship potentially estimates drug dose and drug distribution on post-embolization imaging, enabling identification of potentially under-dosed regions of tumor. The availability of such drug map estimates during TACE could inform treatment decisions, better define treatment endpoints and optimize therapy. __________________________________________________________________________________ National Cancer Institute - Center for Cancer Research Elizabeth Anderson Doctoral Candidate HIV and AIDS Research Abstract removed at request of author Abstract removed at request of author __________________________________________________________________________________ National Cancer Institute - Center for Cancer Research Anuradha Baalsubramanian Postdoctoral Fellow Biochemistry - Proteins Hsp90 of E. coli modulates assembly of FtsZ, the tubulin homolog in E. coli Heat shock protein 90 (Hsp90) is a highly conserved ATP dependent molecular chaperone involved in remodeling, activating and stabilizing numerous client proteins. Since many Hsp90 client proteins have been linked to cancer and other diseases, understanding the functions of Hsp90 is important. The Hsp90 homolog in E. coli, Hsp90Ec, has been shown to cause cell filamentation when overexpressed. By observing the cells under light microscopy, we observed that the filamentous cells had distinct nucleoids, indicating that Hsp90Ec overexpression does not affect chromosomal replication or segregation. To assess if overexpression of Hsp90Ec interferes with the cell division machinery, we tested if FtsZ, a tubulin homolog essential for cell division, assembled into ring-like structures at future sites of cell division as it does in cells not overexpressing Hsp90Ec. We observed by immunofluorescence of fixed cells that FtsZ rings were not detectable in Hsp90Ec overexpressing cells. We also found that FtsZ was present at normal levels in cells overexpressing Hsp90Ec. Together these results suggest high levels of Hsp90Ec affect FtsZ assembly. To test if the Hsp90Ec stabilized negative regulators of FtsZ, we singly deleted genes coding for negative regulators SulA, ClpX, MinC and SlmA but none of these mutants reversed the filamentous phenotype seen in Hsp90Ec overexpressing cells. We next tested the hypothesis that Hsp90Ec prevents FtsZ polymerization. Using purified proteins and fluorescence microcopy, we observed that fluorescently labeled FtsZ formed filaments and bundles in the absence of Hsp90Ec, but not in the presence. Additionally, we showed that light scattering by FtsZ polymers was inhibited when Hsp90Ec was added prior to polymerization. We further observed that an ATP hydrolysis defective Hsp90Ec mutant retained ability to inhibit FtsZ polymerization, consistent with the known ability of Hsp90 to interact with clients independent of ATP hydrolysis. Moreover, we observed that Hsp90Ec client-binding defective mutants exhibited reduced ability to prevent FtsZ polymerization in vitro. In summary, our data show that Hsp90Ec, when overexpressed, inhibits divisome assembly in vivo and prevents FtsZ polymerization in vitro. They suggest that Hsp90Ec may modulate of cell division by interacting and holding FtsZ, possibly slowing cell division during heat stress and other stresses. __________________________________________________________________________________ National Cancer Institute - Center for Cancer Research Defne Bayik Visiting Fellow Immunology - Autoimmune Toll-like receptor 2/1 signaling induces human monocytes to differentiate into immunosuppressive macrophage: implications for the treatment of autoimmune and inflammatory diseases. Autoimmune diseases and chronic inflammatory conditions are characterized by excess immune cell activation. Since inflammatory macrophage have been implicated in the development and progression of such diseases, generation of immunosuppressive macrophages capable of down-regulating pathologic immune responses could provide a therapeutic opportunity. Established methods of generating immunosuppressive macrophage involve stimulation of monocytes with cytokines, which are suboptimal for treatment of chronic conditions. Therefore, we sought alternative methods of immunosuppressive macrophage polarization. Our results demonstrated that Pam3CSK4 (PAM3), an agonist of the Toll-like receptor 2/1 heterodimer (TLR2/1), induces human monocytes to differentiate