Samenvatting
Multiple blood images of stressed and sheared cells, taken by a Lorrca Ektacytometery
microscope, needed a classification for biomedical researchers to assess several treatment options for
blood-related diseases. The study proposes the design of a model capable of classifying these images,
with high accuracy, into healthy Red Blood Cells (RBCs) or Sickle Cells (SCs) images. The performances of
five Deep Learning (DL) models with two different optimizers, namely Adam and Stochastic Gradient Descent
(SGD), were compared. The first three models consisted of 1, 2 and 3 blocks of CNN, respectively, and the
last two models used a transfer learning approach to extract features.
The dataset was first augmented, scaled, and then trained to develop models. The performance
of the models was evaluated by testing on new images and was illustrated by confusion matrices,
performance metrics (accuracy, recall, precision and f1 score), a receiver operating characteristic
(ROC) curve and the area under the curve (AUC) value. The first, second and third models with the
Adam optimizer could not achieve training, validation or testing accuracy above 50%. However, the
second and third models with SGD optimizers showed good loss and accuracy scores during training
and validation, but the testing accuracy did not exceed 51%. The fourth and fifth models used VGG16
and Resnet50 pre-trained models for feature extraction, respectively. VGG16 performed better than
Resnet50, scoring 98% accuracy and an AUC of 0.98 with both optimizers. The study suggests that
transfer learning with the VGG16 model helped to extract features from images for the classification
of healthy RBCs and SCs, thus making a significant difference in performance comparing the first,
second, third and fifth models.
