Abstract:
A deep learning approach to automatic blood cancer diagnosis is proposed in this thesis. The
aim of this research is to develop an efficient system for accurate early-stage diagnosis and
thereby improve healthcare outcomes. The dataset of blood smear images was preprocessed
using techniques such as noise removal, contrast stretching, and normalization to improve the
feature extraction and model training process. Four deep learning models—Xception,
InceptionV3, MobileNet, and ResNet50—were attempted. The best among them was
InceptionV3 with 98%, followed by Xception and MobileNet with 97%. To achieve higher
performance, two hybrid models were attempted: Hybrid Model 1, a combination of Xception,
InceptionV3, and MobileNet, which resulted in 99%, and Hybrid Model 2, a combination of
ResNet50 and VGG16, which resulted in 93%.These results underscore the significance of
model architecture selection and preprocessing for accurate classification. The findings suggest
that AI technology can greatly contribute towards the accuracy and timeliness of the diagnosis
of leukemia, especially in poor-resource environments. The study shows the potential of deep
learning algorithms and, more so, hybrid models for providing accurate, scalable, and efficient
blood cancer detection for the final good of better clinical decision-making and better outcomes
for patients.
