Comparing dropout regularization algorithms for convolutional neural networks identifying malignant cells for diagnosis of leukemia

Abstract: Fast and high quality classifications of cells inflicted with malignant mutations is essential for diagnosing patients with different forms of leukemia, to quickly be able give patients the crucial care they need. Convolutional neural networks (CNNs) can be trained and used for this purpose. This thesis studies CNNs and the application of regularization to create better performing and generalised models, with the purpose of generating highly accurate classifications for nine different forms of malignant white blood cells from the myeloid lineage. This is done to asses what method of dropout regularization is best suited for this type of cell data. To achieve this, three different methods of dropout regularization were studied: Bernoulli dropout; Gaussian dropout; and spatial dropout. This was conducted using a dataset consisting of 106,472 images from 945 patients. The results indicate that models using Gaussian dropout and Bernoulli dropout, respectively, produce the best results, with 87.39\% being the highest accuracy achieved. These two models are also statistically different from a benchmark model not utilizing any form of dropout. This suggests that one of these techniques may be optimal for this type of data. Further studies may be needed to determine which is the best of the two.