J. M. Teague1, D. Socia1, G. An1, S. Badylak2, S. Johnson2, P. Jiang5, Y. Vodovotz2,4, C. Cockrell1 2University Of Pittsburg, McGowan Institute Of Regenerative Medicine, Pittsburgh, PA, USA 4University Of Pittsburg, Department Of Surgery, Pittsburgh, PA, USA 1University Of Vermont College Of Medicine / Fletcher Allen Health Care, Department Of Surgery, Burlington, VT, USA 5Case Western Reserve University School Of Medicine, Center For RNA Science And Therapeutics, Cleveland, OH, USA
Introduction:
The process of wound healing after Volumetric Muscle Loss (VML) injury comprises an intricate balance of multiple functions including inflammation, fibrosis, and cellular proliferation and differentiation. Differential regulation of this balance can induce a range of phenotypes, from fibrotic scar to partial muscle regeneration. Currently, characterization of the functional state of the wound (i.e., gene expression profile in a volume of tissue) requires excision of tissue for subsequent analysis; however, the acquisition of a biopsy itself can reinjure the wound and negatively alter future wound healing dynamics.
In this work, we use a machine-augmented literature review to determine the primary function of 2,749 expressed genes of interest in our canine model. We then trained an artificial intelligence to evaluate images of wound biopsies to regress their functional state.
Methods:
Wound biopsies were obtained by sampling a canine model of VML injury across various locations in the wounds. Molecular analysis determined the RNA expression to identify the 2,749 most variably expressed genes to account for phenotypic differences in healing of VML injury. Various available gene databases were reviewed for each of these 2,749 genes to determine a gene function comprised of four main components with numerous subcomponents, generating a functional profile for each biopsy sample. These profiles were used to label the set of biopsy images. This data was then utilized to train an ensemble neural network using a Metric-Based Semi-Supervised training algorithm.
Results:
The ensemble network regresses detailed functional information on numerous functional categories with <20% error in prediction. Examples of these include: development/ growth inhibition of bone, development/growth of neurite, or activation of neutrophils. These aggregate functions typically made up ~1-10% of the functional activity in the sampled biopsy. The network performance degraded with the functional abundance of associated classes such that genes/functions that were rarely expressed (i.e., on the order of <10-5 of the functional biopsy composition) could not be accurately predicted.
Conclusion:
The project lays the groundwork for further research into the uses of image analysis for many types of medical images. Of specific use to the area of surgery, AIs with the ability to recognize the levels of gene function expression in wounds could be able to determine current healing efficiency, predict wound healing capacity and prognosis, and inform targeted treatment such as smart bandages to treat under of over expression of various wound healing functions. Future experiments would be aimed at the sophistication of this workflow, with hopes that we can enhance the utility of machine-learning approaches for adoption in real clinical practice.