Machine learning on correlation between phase contrast images of 3D bioprinted fibroblasts and cell proliferation

Abstract

Three-dimensional (3D) bioprinting is a field whose popularity has increased worldwide, creating new strategies in many areas, such as organ transplant and biomanufacturing. This approach uses biomaterials such as hydrogels with cells to build layer-by-layer 3D structures.However, some challenges are present as the bioprinted structures are not transparent, translucent, and in some cases, not even homogeneous. This makes the visualization complicated as, using the current optical microscopy techniques, little can be said about the cells inside the gel, requiring immunofluorescence for more detailed information.Fortunately, those issues can be addressed using machine learning approaches, specifically transfer learning, generating a better understanding by extracting, emphasizing, or transforming the available microscopy data. In this in vitro study, we aimed to identify if there is a possible correlation between phase contrast images of 3D bioprinted fibroblasts and the condition of the cells. For that, we created a methodology where we first designed the structure and 3D bioprinted fibroblasts with different printability conditions to obtain distinct viability conditions among the samples. We developed a neural network architecture based on transfer learning of the convolutional base of the VGG16 network and a Random Forest classifier used to classify the pictures.As a result, we could first observe that the UV photocrosslinking duration directly correlated to the samples’ stability. Even mainly without visual distinction between different conditions, the gel without cells appeared more stable during cell culture, and the phase contrast images also showed more defined lines.The fluorescence microscope images presented the DAPI staining of the cells at the border of the gel; over-centrifuged cells had a lower proliferation rate; UV photocrosslinking duration and proliferating cells expressing Ki67 markers are inversely proportional regardless of other printing conditions.The confusion matrices created worked as a visual representation of the predictions. Initially, the network classified the pictures into all the different conditions of printability and obtained an accuracy varying between 15% to 23%. This value increased to 33% to 41% grouping the images per UV photocrosslinking duration, to between 75% to 80% grouping the images as over centrifuged cells, regularly printed cells, and without cells, and finally 85% to 89% grouping the images as gel with our without cells.The 3D model designed present was suitable for machine learning applications and provided first hints on associations between cellular features and visual observations of hydrogels. The neural network was able to classify phase contrast microscopy images depending on how many categories were present, and even though the accuracy was low at classifying multiple categories, decreasing the specificity of the network caused higher sensibility and consecutively higher accuracy showing that there are identifiable correlations between 3D bioprinting properties and phase contrast images.