Abstract:
Detecting cancer early and reliably, especially across different parts of the body, is essential for improving patient care. Yet, many deep learning approaches still focus on a single organ, a single modality, or a tightly controlled dataset. In this thesis, I develop a unified Multi-Cancer Detection System that works across two distinct imaging domains: blood microscopic images and skin dermatoscopic images. The central idea is to use Deep Neural Network (DNN) ensembles to build a system that is not only accurate, but also robust to the variability and complexity that naturally arise in real-world medical data. For blood cancer detection, I designed an ensemble of pre-trained convolutional models whose outputs are combined at the decision level. This ensemble achieved a final validation accuracy of $95.00%$, showing strong reliability in separating clinically important blood cell classes. Skin cancer detection proved more challenging due to severe class imbalance and large variation within lesion types. In this setting, carefully tuned single models consistently plateaued at around $86.89%$ validation accuracy. To address this, I built a dedicated Multi-Model Ensemble System for the dermatoscopic images, which successfully pushed performance beyond this ceiling to $87.83%$ validation accuracy. Taken together, these results show that bringing multiple deep models into a coordinated ensemble can provide more stable and trustworthy predictions than relying on any single network. The dualensemble design across blood and skin suggests a practical path toward a multi-site, imagebased cancer detection tool and offers a flexible foundation that can be extended to additional imaging modalities and cancer types in future work.
