Abstract:
Leukemia, a heterogeneous group of hematologic malignancies, poses significant challenges
in diagnosis and treatment due to its diverse genetic and molecular characteristics. This thesis explores the application of deep learning techniques to classify leukemia subtypes based
on gene expression profiles. Utilizing datasets sourced from the Gene Expression Omnibus
database, this study implements and evaluates three deep learning models: Long Short-Term
Memory, Bidirectional LSTM, and a novel architecture termed xLSTM, which incorporates
custom layers and residual connections for enhanced performance. Data preprocessing involved
standardizing datasets, imputing missing values, and selecting the top 400 features through ChiSquared testing. Models were trained on these processed datasets, and their performance was
assessed through metrics such as accuracy and loss. The xLSTM model demonstrated superior
performance, achieving a final test accuracy of 98.18%, outperforming both LSTM and BiLSTM models. Furthermore, the study examines the integration of multimodal data, combining
gene expression profiles with images data to enhance classification accuracy. Recommendations for researchers emphasize the importance of high-quality, diverse datasets and advanced
preprocessing techniques. Clinicians are encouraged to adopt deep learning tools in practice and
participate in data sharing initiatives, while policymakers are urged to support funding, standardization, and ethical guidelines in AI applications. This research underscores the transformative
potential of deep learning in medical diagnostics, advocating for continued innovation and collaboration across the biomedical and AI research communities to enhance patient outcomes in
leukemia treatment.
