Synthetic Data as a Tool to Combat Racial Bias in Medical AI: Utilizing Generative Models for Optimizing Early Detection of Melanoma in Fitzpatrick Skin Types IV–VI

KVAK, Daniel. Synthetic Data as a Tool to Combat Racial Bias in Medical AI: Utilizing Generative Models for Optimizing Early Detection of Melanoma in Fitzpatrick Skin Types IV–VI. In 3rd International Conference on Medical Image and Computer-Aided Diagnosis (MICAD 2022), Leicester. 2022.

Basic information

Original name

Synthetic Data as a Tool to Combat Racial Bias in Medical AI: Utilizing Generative Models for Optimizing Early Detection of Melanoma in Fitzpatrick Skin Types IV–VI

Name (in English)

Synthetic Data as a Tool to Combat Racial Bias in Medical AI: Utilizing Generative Models for Optimizing Early Detection of Melanoma in Fitzpatrick Skin Types IV–VI

Authors

KVAK, Daniel

Edition

3rd International Conference on Medical Image and Computer-Aided Diagnosis (MICAD 2022), Leicester, 2022

---

Other information

Type of outcome

Prezentace na konferencích

Confidentiality degree

není předmětem státního či obchodního tajemství

References:

URL

Keywords (in Czech)

Autoregressive Models, Computer-Aided Diagnosis, Deep Learning, Generative Adversarial Networks, Melanoma, Synthetic Data, Zero-Shot Learn-ing.

Tags

International impact, Reviewed

---

Abstract

V originále

Assistive tools to aid in skin cancer detection are experiencing an unprecedented rise with the accessibility of robust and accurate deep learning models. However, in the present applications, only a negligible number of dermatology images come from patients with Fitzpatrick skin types IV–VI, representing brown, dark brown or black skin, respectively. In this study, we demonstrate the utilization of Zero-Shot Text-to-Image autoregressive models to generate synthetic medical data for improved balance in training CAD classification models with minimized racial bias. Synthetically generated images of skin lesions were assessed by an experienced dermatologist using the ABCD rule and differential diagnostics, and subsequently validated using a pre-trained ResNet50V2 multi-class classification model.

In English

Assistive tools to aid in skin cancer detection are experiencing an unprecedented rise with the accessibility of robust and accurate deep learning models. However, in the present applications, only a negligible number of dermatology images come from patients with Fitzpatrick skin types IV–VI, representing brown, dark brown or black skin, respectively. In this study, we demonstrate the utilization of Zero-Shot Text-to-Image autoregressive models to generate synthetic medical data for improved balance in training CAD classification models with minimized racial bias. Synthetically generated images of skin lesions were assessed by an experienced dermatologist using the ABCD rule and differential diagnostics, and subsequently validated using a pre-trained ResNet50V2 multi-class classification model.