(D-160) Automated Segmentation of Cellular Autofluorescence Images with a Convolutional Neural Network

Introduction:: NADH and FAD autofluorescence imaging is an emerging non-destructive microscopy technique capable of providing quantitative metabolic information with high spatial and temporal resolution without the need for exogenous stains or dyes. Specifically, cell-specific information on redox state and cofactor protein binding can be extracted from maps of the optical redox ratio of FAD/(NADH+FAD) autofluorescence intensity and NADH fluorescence lifetime imaging (FLIM), respectively. These quantitative measures have been used to understand cell metabolism in various fields, including cancer research and regenerative medicine. However, quantitative analysis of cell redox state or FLIM requires accurate segmentation of cellular regions of interest (ROI). Masks defining the ROI are typically generated by manually tracing the region of interest, which is time-consuming and subjective. Masks defined based on intensity thresholds are also often inaccurate and/or lack broad applicability across cell types or acquisition settings. Convolutional neural networks (CNNs) could automate cell segmentation because of their extraordinary ability to detect features and classify images. The goal of this study is to train a convolutional neural network to segment a range of cell types from different acquisition settings (traditional fluorescence intensity images and FLIM images).

Materials and Methods:: NADH and FAD autofluorescence images were summed from 181 fields of view using a mixture of existing data from mesenchymal stem cells, 3T3 fibroblasts, and keratinocytes. An additional 154 NADH FLIM images collected through time-correlated single photon counting were added to the data set. Image intensities were normalized to range between 0 and 1 and were manually traced to create ground truth logical masks. A convolutional neural network was designed and trained in MATLAB based on a U-net architecture with an encoder depth of 4, consisting of convolution blocks of 64, 128, 256, and 512 feature detectors. Training occurred over 50 epochs with data randomly partitioned into training (80%), validation (10%), and testing (10%) sets.

Results, Conclusions, and Discussions::

The trained CNN segmented images in the independent testing set with 85% accuracy with comparable segmentation accuracy across both typical autofluorescence intensity images and FLIM data (see Figure). Sensitivity to cellular ROIs was particularly high (91%), but a lower specificity (75%) was the product of the network defining some small regions of autofluorescence outside of the cell bodies as part of the ROI (see Figure). In the near-term, the network will be retrained on a larger dataset. To improve specificity, we will apply a Gaussian filter to CNN’s probability matrix to generate smoother masks and eliminate small objects. We will also evaluate the cell classification threshold through a receiver operating characteristic curve to further improve accuracy.
We created a CNN capable of differentiating between cells and background from a broad range of autofluorescence images. The CNN demonstrated promising overall accuracy, which will be further refined. In future work, the trained network will enable fully automated quantitative analysis of cellular redox state and cofactor binding status by eliminating the need for subjective and time-consuming manual segmentation

Acknowledgements (Optional): :

References (Optional): :