Kidney tumor detection using deep learning and dimensional reduction-assisted machine learning methods

This thesis proposes a deep-learning and dimensionality-reduction–assisted machine learning approach for fast and accurate detection of renal tumors from CT images. Two large CT datasets (V1 and V2) were preprocessed; texture features were extracted using the Gray Level Co-occurrence Matrix (GLCM) and reduced to two dimensions with t-SNE. The resulting low-dimensional features were used to train KNN, Bagged Trees, Decision Tree, Fine Gaussian SVM, and a three-layer Artificial Neural Network. For comparison, deep learning experiments were conducted with AlexNet, EfficientNet-B0, Darknet-53, Xception, and DenseNet-201. Results show that DenseNet-201 achieved 99.75% accuracy at a 0.0001 learning rate, while the proposed cascaded GLCM + t-SNE pipeline delivered 99.65% accuracy on V1 and 99.98% on V2 with KNN, alongside markedly smaller model size and higher prediction speed. Using t-SNE provided substantial speed and memory gains at accuracy comparable to GLCM-only features. The method strengthens clinical decision support thanks to low system requirements, real-time applicability, and robust, dataset-agnostic performance. Evaluation relied on accuracy, precision, recall, F1, and specificity metrics derived from confusion matrices, and hyperparameters were selected via grid search. Improvements in model size and inference speed facilitate deployment on memory-constrained edge devices and enhance overall reliability. Future work will explore additional imaging modalities, transfer learning, and explainable AI to improve generalizability and interpretability.