Enhancing Leukemia Diagnosis with Synthetic Data and Explainable Deep Learning Architectures

Malik, Ibrahim (2025) Enhancing Leukemia Diagnosis with Synthetic Data and Explainable Deep Learning Architectures. Masters thesis, Dublin, National College of Ireland.

Preview	PDF (Master of Science) Download (2MB) | Preview
Preview	PDF (Configuration Manual) Download (213kB) | Preview

Abstract

Leukemia diagnosis through microscopic blood smear analysis remains time-intensive, error-prone, and dependent on expert interpretation. To solve issues of interpretability and data scarcity, this study introduces a first comprehensive framework combining Generative Adversarial Networks (GANs), Convolutional Neural Networks (CNNs), Vision Transformers (ViTs), and Explainable AI (XAI).

For artificial data augmentation, three GAN variations (DCGAN, WGAN, and cGAN) were used; conditional GANs produced better image quality (FID: 1.1897, SSIM: 0.8869). Using the CNMC dataset, classification showed different architectures responding differently to simulated data. Conventional data quality assumptions were challenged when ViTs showed higher results with WGAN augmentation, whereas CNNs performed ideally with cGAN augmentation. The best ViT-WGAN configuration achieved 74% accuracy, presenting a 167% gain crucial for reducing missed diagnoses, with significant sensitivity improvements from 0.187 to 0.500.

CNN-ViT hybrid architectures are seen to perform worse than individual models, indicating that convolutional and attention methods do not represent features incompatibility. Explainable AI analysis using Grad-CAM, LIME, and SHAP shows models learned clinically significant morphological characteristics, with focus on cell borders and dispersed properties reflecting pathologist decision patterns.

The framework addresses class imbalance through data generation while providing interpretable predictions essential for adoption. Findings show promise for clinical application, providing pathologists with transparent and precise diagnostic support tackling issues of confidence and performance in medical AI application.

Item Type:	Thesis (Masters)
Supervisors:	Name Email Agarwal, Bharat UNSPECIFIED
Subjects:	Q Science > QH Natural history > QH301 Biology > Methods of research. Technique. Experimental biology > Data processing. Bioinformatics > Artificial intelligence Q Science > Q Science (General) > Self-organizing systems. Conscious automata > Artificial intelligence Q Science > Life sciences > Medical sciences > Pathology > Tumors > Cancer R Medicine > Healthcare Industry
Divisions:	School of Computing > Master of Science in Data Analytics
Depositing User:	Ciara O'Brien
Date Deposited:	01 Jul 2026 11:40
Last Modified:	01 Jul 2026 11:40
URI:	https://norma.ncirl.ie/id/eprint/9437

Actions (login required)

View Item