The Effect of the Splitting Protocol on Performance in Dermatoscopic Skin Lesion Classification: A Controlled Measurement of Data Leakage
The 4th International Conference on Applied Mathematics in Engineering (ICAME’26), Balıkesir, Türkiye, 2 - 04 Eylül 2026, ss.155, (Özet Bildiri)
- Yayın Türü: Bildiri / Özet Bildiri
- Basıldığı Şehir: Balıkesir
- Basıldığı Ülke: Türkiye
- Sayfa Sayıları: ss.155
- Açık Arşiv Koleksiyonu: AVESİS Açık Erişim Koleksiyonu
- Manisa Celal Bayar Üniversitesi Adresli: Evet
Özet
Skin cancer is one of the most common cancers in the world, and early diagnosis has a direct effect on treatment
success. Classifying skin lesions automatically from dermatoscopy images is an active research topic for clinical
decision support systems. In this study, four deep learning architectures (DenseNet121, ConvNeXt-Tiny,
ResNet50, and EfficientNet-B3) were tested on the HAM10000 dataset [1] for seven-class skin lesion
classification. The models were compared in two scenarios. In the first scenario, the data was split at the lesion
level, so images of the same lesion could not appear in more than one set. In the second scenario, data augmentation
was applied before the split. This was done on purpose to create data leakage, and the same four models were
retrained under this protocol. In both scenarios, class imbalance was handled with undersampling and
WeightedRandomSampler. Albumentations was used for data augmentation [2], and OneCycleLR was used as the
learning rate scheduler during training [3]. Each experiment was repeated with three random seeds (13, 21, 42),
giving 24 trained models in total. The results are reported as mean and standard deviation. In the first scenario,
DenseNet121 reached the highest test accuracy (77.39% ± 1.00), but the gaps between the models stayed within
the standard deviations. In the second scenario, the test accuracy of every model went up. The mean test accuracy
rose from 76.47% to 83.70%, an increase of 7.23 points. For DenseNet121, it rose from 77.39% to 86.95%. The
size of the increase was different for each architecture (from 2.77 to 9.57 points), but it was positive in all twelve
runs. This finding shows why a lesion-level split is necessary in medical imaging studies.