AI-driven diagnostics: accuracy vs. ethical considerations

Abstract

Artificial intelligence (AI) is transforming the work of a healthcare provider, providing physicians with an opportunity to diagnose and treat patients differently. This thesis explores the application of deep learning in medical imaging in a responsible way, using as a case study the brain-tumour segmentation. To enhance accuracy and reduce bias, a U-Net model was trained using the Brain Tumor MRI dataset (Nickparvar, 2021) that is normalized, augmented, and crossvalued. This model obtained a Dice score and an IoU of 0.961 and 0.937, respectively, which indicated that the model segmented the tumours effectively and reliably. Although the results were strong, the real challenge was to keep the system within ethical and legal limits during its design.

All patient data were anonymized before training, and checks were performed to ensure that the model treated each case fairly. To comply with GDPR requirements, audit trails were implemented so that every experiment could be traced if necessary. These measures went beyond administrative compliance, demonstrating that trust in medical AI is rooted in careful and transparent practice. Ultimately, accuracy alone does not make a system responsible; what matters is that the technology is developed fairly and openly, and that it respects the individuals whose data make the research possible.