The Impact of Data Leakage on Brain Tumor Classification Beyin Tümörü Siniflandirmasinda Veri Sizintisinin Etkisi
34th Signal Processing and Communications Applications Conference, SIU 2026, İstanbul, Turkey, 7 - 10 July 2026, (Full Text)
- Publication Type: Conference Paper / Full Text
- Doi Number: 10.1109/siu71813.2026.11636504
- City: İstanbul
- Country: Turkey
- Keywords: 5-fold cross-validation, Brain tumor classification, data leakage, deep learning, wrong split
- Yıldız Technical University Affiliated: Yes
Abstract
Automatic classification of brain tumors from Magnetic Resonance Imaging (MRI) is important for early diagnosis and treatment planning. However, in widely used public datasets, improper data splitting strategies can lead to data leakage, resulting in performance estimates that are higher than the true generalization ability. Although this risk has been generally discussed in the literature, studies that systematically compare patient-independent splitting and slice-based splitting on the Figshare-based brain tumor dataset within the same experimental framework remain limited. In this study, eight deep learning models are evaluated on a Figshare-based dataset using 5-fold cross-validation. The findings show that patient-based independent splitting provides more realistic and reliable results, whereas slice-based splitting can overestimate the true generalization performance of the model.