Deep learning-based analysis and classification of peripheral blood smear results using image enhancement techniques

In this thesis, the impact of image enhancement techniques applied to peripheral blood smear images on classification performance was comprehensively analyzed, and the obtained findings were evaluated within the framework of a deep learning-based hybrid model. The primary objective of the study is to develop a holistic deep learning architecture that enhances diagnostic accuracy in microscopic images by integrating preprocessing steps that improve visual quality, attention mechanisms, and texture-based feature extraction. In line with this, the proposed CNN model was designed to provide both rich feature-level representation and high decision-level reliability. In the initial phase, transfer learning was applied using 19 different pre-trained CNN architectures. Although VGG16 yielded relatively lower initial accuracy, it was chosen as the core model for the experimental process due to its structural simplicity, modularity, and high potential for enhancement. Likewise, three publicly available white blood cell datasets were evaluated, and "PBC Dataset Normal DIB" was selected based on class balance, image quality, and diversity. The hypothesis posits that image enhancement, optimized illumination levels, attention module integration, texture feature extraction, and ensemble architectures would each individually and in combination significantly improve classification performance. Accordingly, the VGG16 model was trained under nine different brightness levels ranging from -80 to +80, and +40 illumination was found to yield the best performance compared to original images. Subsequently, CBAM (Convolutional Block Attention Module) was integrated into the VGG16 model to enhance attention to relevant image regions, while GLCM and LBP-based texture features were extracted to provide multidimensional input representation. Three complementary CNNs, MobileNetV2 (lightweight and fast), ResNet50 (deep and robust), and EfficientNet-B0 (balanced and efficient) were fused into an ensemble architecture. The resulting ESA model achieved 99.82% classification accuracy, and its generalizability and stability were statistically validated using k-fold cross-validation.