<![CDATA[Skin disease classification with deep learning has shown promising performance, however many models are primarily trained on datasets featuring light skin tones, which raises question about their effectiveness across a variety of akin types. This study analyses the response of a ResNet50 model based on transfer learning when faced with different skin tones in order to classifying skin disease using medical images. The model was trained on the HAM100000 which categorized into three classes: benign, malignant, and non-neoplastic. A bias analysis was then performed using the Fitzpatrick 17k dataset. The model demonstrated an overall accuracy of 70.85%, a precision rate of 74.03%, and a recall rate of 65.51%. Further analysis showed that the model had a consistent pattern of predicting malignant cases, which increased with darker skin tones, rising from 54% to 68.3%. To mitigate this issue, a threshold tuning approach was applied. After mitigation, the model achieved an accuracy of 74%, a weighted F1-score of 76%, dan a macro F1-score of 55%. Fairness evaluation after mitigation showed tha the proportion of malignant predictions increased from 56,3% in FST I to 69,9% in FST VI. These findings suggest that threshold tuning can improve classification performance and partially reduce bias intensity.]]>
Copyrights © 2026
