ASSESSING PREDICTION OUTCOMES OF CONVOLUTIONAL NEURAL NETWORKS TRAINED ON PARASITE AND HEALTHY CELL IMAGES FOR SICKLE-CELL AND MALARIA

To address the implications of using a CNN, a machine learning model was trained with parasite (positive) and healthy (negative) cell images of each respective disease, sickle-cell and malaria. In a world where the field of AI-assisted diagnostics has experienced exponential growth in the past decade, the need for significant research is necessary. Through this project, the goal is to have two working CNN models (one model per one disease) that can accurately predict if a human has sickle-cell or malaria based on positive and negative cell image data sets. The initial hypothesis is as follows, if a CNN model is trained on positive and negative images for both diseases, it can accurately predict the presence of both with 90% accuracy. Both datasets came from Kaggle, but the Malaria data set originated from NIH datasets and the sickle-cell data set images were produced from a research study located in Uganda, Africa. Preprocessing steps included resizing images to a standard 255 by 255 pixels and color standardization to ensure the model is not skewed based on unintended biases. The model was mainly trained through a random forest algorithm, a supervised learning mechanism, and was split into a 10% testing set, 80% training set, and 10% validation set. These proportions differed when running ablation tests, but this was the standard testing split done. Evaluation techniques include detailed confusion matrixes, accuracy curves, and loss curves to interpret this model’s results extensively.