Hybrid deep learning approaches for the multi class medical image classification

This thesis presents a hybrid framework that combines deep learning architectures, meta-heuristic approaches, dimensionality reduction methods, and tree-based classification algorithms for the classification of blood cells with different characteristics. The study incorporates deep learning architectures such as Residual Network (ResNet), including ResNet18 and ResNet34, meta-heuristic methods such as covariance matrix adaptation evolution strategy (CMA-ES) and particle swarm optimization (PSO), dimensionality reduction methods like principal component analysis (PCA), independent component analysis (ICA), uniform manifold approximation and projection (UMAP), and t-distributed stochastic neighbor embedding (t-SNE), along with tree-based algorithms like Random Forest (RF), eXtreme gradient boosting (XGBoost), and light gradient-boosting machine (LightGBM). Research has shown that deep learning architectures such as ResNet18 and ResNet34 are highly effective in image processing tasks. Furthermore, literature studies indicate that tree-based algorithms like RF, XGBoost, and LightGBM achieve very high classification performances. Building on the success of these approaches, the ResNet18 and ResNet34 models are used for feature extraction, while tree-based algorithms such as decision trees are employed for classification. Additionally, past literature demonstrates that appropriate feature selection and/or dimensionality reduction approaches are important in mitigating overfitting, enhancing generalization performance, and/or reducing computational complexity before classification. In this context, after the pre-training stage of the deep learning architecture, meta-heuristic algorithms like CMA-ES and PSO are used for feature selection, and dimensionality reduction algorithms such as PCA, ICA, UMAP, and t-SNE are used in different combinations to investigate their effects on classification performance. The performance of the proposed hybrid methods is evaluated using the BloodMNIST dataset, which contains eight different types of blood cells from the MedMNIST dataset. Acknowledging that the performance of machine learning methods heavily depends on hyperparameter settings, specific hyperparameters were selected for the ResNet architecture, RF, XGBoost, LightGBM classifiers, and CMA-ES and PSO meta-heuristics, and the best values for these parameters were determined. In this thesis, the performance of hybrid methods, incorporating different combinations of the selected approaches, was thoroughly examined to determine the most suitable hybrid framework. In this context, experimental studies were conducted on four different hybrid methods. The first hybrid framework involved extracting features from the BloodMNIST dataset using ResNet architectures during the pre-training stage, followed by investigating the classification performance of RF, XGBoost, and LightGBM algorithms using these extracted features. In this examination, it was observed that the LightGBM classifier outperformed the others. The second hybrid framework investigated the effects of adding dimensionality reduction methods between the ResNet and classifier layers to reduce computational complexity and/or remove outliers to improve generalization performance. The analysis showed that PCA and ICA algorithms improved performance, while UMAP and t-SNE algorithms decreased it. In the third hybrid framework, meta-heuristic approaches were applied between the ResNet and classifier layers for feature selection to obtain more refined features, and their effects on performance were examined. The analysis revealed that meta-heuristic algorithms, particularly after hyperparameter optimization, improved performance. Among CMA-ES and PSO methods, the highest performance was achieved with the PSO algorithm. In the fourth hybrid framework, meta-heuristic algorithms were combined with dimensionality reduction algorithms between the meta-heuristic and classifier layers to further enhance performance. The analysis showed that using meta-heuristic and dimensionality reduction algorithms together decreased performance, while selecting the best features using meta-heuristic approaches improved classification performance. This thesis demonstrates that hybrid use of feature extraction through ResNet architectures, feature selection with meta-heuristic approaches such as PSO, and classification with tree-based approaches such as LightGBM improves classification performance. The hybrid framework offers a structure that can incorporate various deep learning architectures, different meta-heuristic methods for feature selection, diverse dimensionality reduction approaches, and different classification methods.