Towards Ethical Dermatology: Mitigating Bias in Skin Condition Classification

Anshul Pundhir, Sanchit Verma, Balasubramanian Raman

Abstract: Deep neural networks have proven to be highly effective in efficiently handling various medical tasks, including classification and segmentation in the healthcare domain. Specifically, convolutional neural networks (CNNs) have gained significant popularity in aiding dermatologists with the diagnosis of skin lesions. The CNNs are found to surpass the performance of dermatologists and reduce the need for manual intervention. However, a potential drawback of deep neural networks is their susceptibility to biased predictions towards minority subgroups within the dataset they are trained on, as they effectively learn from the underlying data distribution. In dermatology datasets, researchers have highlighted the negative impact caused by the under-representation of individuals with dark skin tones, which leads to social and medical discrimination. This data imbalance poses challenges when deploying deep neural networks on a larger scale, as it can result in unfair outcomes and exacerbate existing biases. To address this issue, we utilize a debiasing approach based on variational autoencoders that effectively mitigate the data bias present in deep neural networks. To evaluate the effectiveness of our proposed approach, we conducted experiments using a recent benchmark dataset known as Fitzpatrick-17k. This dataset consists of clinical images representing various skin conditions, each labeled for the disease class and Fitzpatrick skin tone. By leveraging the inherent latent data distribution across different classes, we demonstrated significant improvements in mitigating the bias related to skin tone within the dermatology dataset. As a result, we observed an enhanced classification rate for the under-represented minority classes. To validate our proposed approach, we employed both quantitative and qualitative measures. Our model not only outperformed existing benchmark approaches for skin condition classification but also effectively addressed the detrimental effects caused by inherent bias in the classifier. Furthermore, we employed the Uniform Manifold Approximation and Projection (UMAP) technique to visually showcase the robust learning of the data distribution by our model, providing evidence of its generalization capabilities. We will share our source code publicly for reproducibility to facilitate further research and validation of our approach.