Particle Detection in Bone Marrow Using CNNs and Transformer Networks

Sammanfattning: The critical role of hematological stem cells, concentrated in bone marrow, in disease diagnosis makes the detection of these cells an imperative task. The current process, however, is slow and heavily reliant on expert interpretation, underscoring the need for automation. In this thesis, we collected a 20x microscopy dataset of real and fake bone marrow particles and fine-tuned several CNNs and Transformer networks on the task of classifying them. We performed a hyperparameter investigation, explored whether illuminating the sample from an angle improves performance and analysed what the networks are looking for with several interpretability methods. The best results were obtained by fine-tuning a Swin Transformer V2S, using images captured standard illumination settings. This model reached 97.19% on the test set. Different hyperparameters were evaluated with the result that most hyperparameters are of little significance, apart from dataset size where the accuracy increased logarithmically with the dataset size. A much faster model was made by quantising ResNet18 to 8-bits, which had an accuracy of 96.75% while only taking 59ms per image for inference on microscope hardware, making it feasible to use in an automated bone marrow examination process in the future.