A Comparison of Image Recognition Algorithms for Cell Phenotyping in Optofluidic Time-Stretch Microscopy

Wanyue Zhao,Cheng Lei,Honghao Huang,Keisuke Goda,Hongwei Chen
DOI: https://doi.org/10.1117/12.2537782
2019-01-01
Abstract:Optofluidic time-stretch imaging system has enabled high-throughput phenotyping of cells with unprecedented high speed and resolution. However, significant amount of raw image data is produced, which requires recognition algorithm with not only high accuracy but also high speed to analyze image data efficiently. In this paper, we compare the performance of popular feature extraction methods and learning-based classification algorithms on time-stretch microscopy image recognition. The applied image recognition system comprises an outlier detection step, feature extraction method and classification. The main concept of outlier detection uses DBSCAN (Density-Based Spatial Clustering of Applications with Noise) to eliminate error images. Gabor wavelet, HOG (Histograms of Oriented Gradients), LBP (Local Binary Pattern) and PCA (Principal Components Analysis) are applied and compared as the feature extraction methods. Finally, with a set of extracted features, the computing time and accuracy of SVM (Support Vector Machines), LR (Logistics Regression), ResNet (Residual Neural Network) and XGBoost (Extreme Gradient Boosting) classification algorithms are evaluated. The tested cell image datasets are acquired from high-throughput imaging of numerous drug-treated and untreated cells (N = similar to 21,000) with an optofluidic time-stretch microscope. Results show that PCA feature extraction and XGBoost classification proves to be the fastest algorithms with the highest level of accuracy. DBSCAN outlier detection helps to improve the recognition accuracy by 2% approximately. Therefore, we propose a recognition algorithm consisting of DBSCAN outlier detection, PCA feature extraction and XGBoost classification as a promising solution to process the image data of high-throughput optofluidic time-stretch microscopy accurately and rapidly.
What problem does this paper attempt to address?