Limited Time Offer

Claim a 25% discount on all eLearning courses (including credentials) with code ELEARN25.

Offer is valid from March 10-31. Public courses excluded from promo. 

(89f) Capturing Biomaterial Degradation and Tissue Formation Following Implantation of Silk-ECM Composite Sponges Using Machine Learning and Kinetic Modeling

Stoppel, W. - Presenter, University of Florida
Jameson, J. F., University of Florida
Butler, J., University of Florida
Peeples, J., University of Florida
Kotta, N., University of Florida
Biomaterials available for surgeons that are applicable for large soft tissue injuries can range from natural to synthetic materials, with a wide variety of positives and negatives depending on the clinical application. Silk fibroin, a protein extracted from Bombyx mori silkworm cocoons, has been utilized for cosmetic surgeries and gained Food and Drug Administration 510(k) clearance for some applications (for example via Sofregen, Inc., Boston, MA). We are specifically interested in quantitatively describing how silk-based sponge-like materials degrade and support neo-tissue formation as a function of the parameters we can control when making the implantable sponge material. Current material design is a bit of a guess and check experience and thus we aim to streamline material formulation decisions through predictive modeling and acceleration of quantitative analysis post implantation through machine learning-based approaches to histological image analysis.

We have developed a kinetic model to describe silk scaffold degradation, determining that the enzyme and the enzyme concentration are major drivers of the silk sponge degradation. We have also shown that addition of extracellular matrix proteins and bioactive molecules in low concentrations that are sufficient to elicit changes in cell infiltration are not sufficient to elicit changes in degradation kinetics in vitro. To better understand how in vitro data correlate with in vivo observations, we needed tools to analyze cell infiltration and rates and scaffold degradation spatiotemporally and quantitatively. Together with the Zare Lab, we proposed a convolutional neural network (CNN) model with novel spatial histogram layer(s) that can more accurately identify and segment regions of adipose tissue in images of hematoxylin and eosin or Masson’s Trichrome stained sections. Use of the CNN model allowed for determination of the optimal formulation for the biomaterial that, in this case, limited adipose tissue formation compared to other scaffold formulations. We compared our proposed method, Jointly Optimized Spatial Histogram U-Net Architecture (JOSHUA), to the baseline U-Net model currently used in biomedical image segmentation as well as to a version of both models with a supplemental attention mechanism (JOSHUA+ and U-Net+). The inclusion of histogram layer(s) in our models indicates improved performance through qualitative and quantitative (dice coefficient, intersection over union) evaluation. These efforts represent two new tools for studying cell infiltration, tissue formation, and scaffold degradation within silk sponges.