Deep learning cell segmentation in chondrocyte viability assessment using nonlinear optical microscopy

In the recent studies of cartilage imaging with nonlinear optical microscopy, we discovered that autofluorescence of chondrocytes provided useful information for the viability assessment of articular cartilage. However, one of the hurdles to apply this technology in research or clinical applications is the lack of image processing tools that can perform automated and cell-based analysis. In this report, we present our recent effort in the cell segmentation using deep learning algorithms with the second harmonic generation images. Two traditional segmentation methods, adaptive threshold, and watershed, were used to compare the outcomes of different methods. We found that deep learning algorithms did not show a significant advantage over the traditional methods. Once the cellular area is determined, the viability index is calculated as the intensity ratio between two autofluorescence channels in the cellular area. We found the viability index correlated well with the chondrocyte viability. Again, deep learning segmentation did not show a significant difference from the traditional segmentation methods in terms of the correlation.