Integrated Machine Learning Techniques for Enhanced Lymphoma Classification

Abstract

Recent advancements in digital pathology have underscored the critical role of automated medical image analysis for disease diagnosis. While texture-based approaches are widely employed, conventional handcrafted feature extraction methods often yield suboptimal results due to high inter-class correlations in medical imagery. This thesis proposes a novel hybrid framework that synergizes Local Phase Quantization (LPQ) with chaotic-weighted Gaussian filtering and Bat Algorithm optimization to enhance lymphoma subtype classification. Our methodology begins with a deep texture analysis of histopathological images using an adaptive filter bank, where chaotic systems dynamically weight filter parameters. The extracted descriptors are processed via LPQ for robust feature representation, while the Bat Algorithm optimizes filter configurations to maximize discriminative power. Validated on a multicenter lymphoma dataset with inherent staining variability, the system achieves 96.17% Accuracy, surpassing state of the art handcrafted techniques.