Machine Learning Approaches in Medical Diagnosis

Abstract

Brain tumors are among the most life-threatening neurological conditions, characterized by abnormal and uncontrolled cell growth within brain tissues. Early identification of tumor type and stage is crucial for effective intervention, yet conventional diagnostic methods often struggle to detect tumors in their early or subtle stages. This study proposes advanced deep-learning models for multi-class classification of brain tumors and machine-learning approaches for predicting chemical properties of anti-cancer drugs to support new drugs development. A hybrid deep-learning model was developed for classifying three tumor categories glioma, Meningioma and pituitary tumors using MRI scans. the model was trained on the brain tumor MRI dataset and integrates ResNet50 and Efficient Net50 as its base ar-chitectures, combined through a novel triangular dense-layer fusion strategy to optimize multi-stage feature extraction. The proposed model achieved a test accuracy of 97%, recall of 97%, and an AUC score of 99%, demonstrating its effectiveness for early and accurate brain-tumor classification. In second phase of this research, machine-learning algorithms and topological indices were used to analyze molecular structure of thirteen anti-cancer drugs. Topological in-dices were computed using a python program, and true physicochemical properties were retrieved from the ChemSpider database via an automated script. QSPR and SHAP anal-ysis were performed to identify indices most predictive of each physicochemical property. Machine-learning models were then trained on these features to develop generalized pre-dictive models for key durg properties, providing a foundation for computationally assisted drug discovery. Overall, this thesis demonstrates the combined potential of deep learning for early tu-mor detection and machine-learning for anti-cancer drug analysis, ultimately contributing to improved diagnostic accuracy and future therapeutic development