Detection of Brain Cells in Optical Microscopy Based on Textural Features with Machine Learning Methods

The problem of detecting neurons in optical microscopy is considered by the example of Nissl-stained mouse brain slices. The proposed algorithm consists of the following steps: preprocessing, textural feature extraction, pixel classification, and pixel clustering. When solving this problem, we investigate various preprocessing methods, machine learning algorithms, and textural features. At the classification step, the k nearest neighbor (kNN) algorithm or the naive Bayes classifier (NBC) is used to determine whether each pixel of the image belongs to the neuron’s cell body. In this paper, we investigate textural features of two types: features based on the normalized histogram and features based on the gray level co-occurrence matrix (GLCM). To find the centers of the neurons, all pixels that belong to the neurons are clustered using the mean shift algorithm. It is shown that the best detection quality (precision = 0.82, recall = 0.92, and F1 = 0.86) is achieved with GLCM-based features and neighborhood radius R = 6. It is also shown that the selection of different preprocessing algorithms significantly affects the detection result. In terms of detection quality, the kNN algorithm outperforms the NBC. On the dataset used, the selection of the parameter k > 15 does not significantly improve the quality of detection. The proposed method yields the result similar to that achieved in [1]: Recall(A) = 865%. In sampling tests on some microscopy images from the Broad Bioimage Benchmark Collection (BBBC), the proposed approach shows the best or equivalent quality in detecting the number of cells on the image. For detection, the algorithm uses only local textural features, which removes restrictions on the parallelization of computations.