Evaluation of Image Classification for Quantifying Mitochondrial Morphology Using Deep Learning

Abstract

Background: Mitochondrial morphology reversibly changes between fission and fusion. As these changes (mitochondrial dynamics) reflect the cellular condition, they are one of the simplest indicators of cell state and predictors of cell fate. However, it is currently difficult to classify them using a simple and objective method.

Objective: The present study aimed to evaluate mitochondrial morphology using Deep Learning (DL) technique.

Methods: Mitochondrial images stained by MitoTracker were acquired from HeLa and MC3T3-E1 cells using fluorescent microscopy and visually classified into four groups based on fission or fusion. The intra- and inter-rater reliabilities for visual classification were excellent [(ICC(1,3), 0.961 for rater 1; and 0.981 for rater 2) and ICC(1,3), respectively]. The images were divided into test and train images, and a 50-layer ResNet CNN architecture (ResNet-50) using MATLAB software was used to train the images. The datasets were trained five times based on five-fold cross-validation.

Result: The mean of the overall accuracy for classifying mitochondrial morphology was 0.73±0.10 in HeLa. For the classification of mixed images containing two types of cell lines, the overall accuracy using mixed images of both cell lines for training was higher (0.74±0.01) than that using different cell lines for training.

Conclusion: We developed a classifier to categorize mitochondrial morphology using DL.

Keywords: Mitochondrial morphology; ResNet; deep learning; fission; fusion; mitochondrial dynamics.