Hybrid RBC Morphology Analysis and Diagnostic Framework for β- Thalassemia Using SEBlock-CBAM Enhanced MobileNetV2, TabNet with Optuna Optimization and SMOTE-ENN Resampling

This research presents a multidimensional, innovative approach for diagnosing β-thalassemia, converging discovers from multiple modalities such as haematological parameters and images of RBC morphology and applying deep learning techniques. A combination of MobileNetV2 architecture, Squeeze-and-Excitation (SEBlock), and Convolutional Block Attention Modules (CBAM) was coupled with the image analysis functionality, while the TabNetClassifier, trained optimally using Optuna, targeted the numerical data functionalities. Advanced preprocessing techniques included SMOTE-ENN resampling and logarithmic transformation to address class imbalances and skewed data distributions. Furthermore, the proposed approach achieved a precision of 0.90, recall of 0.892, and F1-score of 0.893 for β- thalassemia classification, outperforming traditional diagnostic methods. The RBC classification model demonstrated an average accuracy of 90.8%, with a Cohen's Kappa score of 0.8861, indicating strong agreement between predicted and actual classifications. LIME tool was explored to provide clinicians with actionable insights. However, the study is limited by its reliance on a single external dataset for validation. Future work will focus on multi-center trials to ensure robustness across diverse populations.