Segmentation of Neuronal Cells Using Simplistic Methods : A Comparison of the Mean Shift Algorithm and Otsu’s Method

Abstract: Information regarding specific neuronal characteristics, such as shape and distribution, is essential for quantifying the brain structure and modelling accurate computer simulations. To this end, it is important to perform cell segmentation; to isolate the cells in a given image from the surrounding tissue, so it can be further analysed. While many state-of-the-art approaches involving machine learning are complex, classical algorithms still see frequent use due to their simplicity. This research paper evaluated the applicability of both the Mean Shift Algorithm and Otsu’s Method; two classical algorithms with high potential which have not seen much use. The algorithms were applied on a representative set of images from a specific experiment in the Allen Mouse Brain Atlas and evaluated based on true and false pixel identifications. The Mean Shift Algorithm was shown to consistently undersegment the existing neurons, but also rarely introduced false positives, leading to a higher reliability of its segmentation. On the other hand, Otsu’s Method was found to be generally more accurate and it detected almost all of the existing neurons, but was prone to oversegmentation and often incorrectly classified other tissue as neurons. It was also less consistent, more likely to produce anomalous results, and more negatively affected by image artefacts than the Mean Shift Algorithm, which was largely unaffected. Furthermore, both algorithms were found to decrease in accuracy as the cell density increased. The applicability of either algorithm depends on the specific use case, but they are mostly relevant to small scale use where accuracy is less important and complexity is unwarranted, such as in early prototyping stages. While the Mean Shift Algorithm was generally less accurate, it was more reliable, indicating that it can be used for isolating regions which are almost guaranteed to be neurons. Otsu’s Method could potentially be used as a masking filter in order to remove unimportant tissue before applying additional methods, and consequently reducing the input size and improving performance.