Seeing is Believing: How AI might transform skin cancer education

The integration of digital technologies, e.g., smartphone apps, has been shown to represent an impactful advancement in training for melanoma diagnosis.1

As artificial intelligence (AI) rapidly advances, its integration into dermatology, particularly through Cycle-Consistent Adversarial Networks2 (Cycle-GANs), we present possible new horizons in patient education and skin cancer diagnosis. A GAN functions with two neural networks: a generator creating images and a discriminator evaluating them, producing increasingly realistic results. A Cycle-GAN links two GANs for bidirectional data transformation between domains. GANs have already found a significant application in dermatology, particularly for data augmentation in skin cancer classification models. The GANs have been enhancing the robustness and accuracy of these diagnostic models.3 We propose the application of Cycle-GANs in the transformation of benign dermoscopic images into artificially created malignant counterparts. This technology allows dermatologists to visually demonstrate to patients the subtle differences between benign and malignant lesions using actual dermoscopic images from the patient's body. During skin screening procedures, dermatologists can present both the original benign image and the AI-generated malignant version. This side-by-side comparison aids in explaining why certain lesions do not require excision and what changes to look for in the future, potentially making patients more attentive. Conversely, for lesions appearing malignant, the AI could generate a benign counterpart to illustrate the necessity of excision.

Case Presentation

Figure 1 shows a dermoscopic image of a benign lesion from the ISIC dataset4 and an artificially created malignant counterpart as well as selected frames from a frame interpolation, showing the gradual transformation from the original benign image to the artificial malignancy image.

 Figure 1 Top: Original benign skin lesion from ISIC dataset. Centre: Selected frames from the frame interpolation video generated with Runway (2024 Runway AI, Inc., New York, USA).  Bottom: AI-generated image showing the lesion's potential malignant progression.

Utilizing frame interpolation, we create a seamless and gradual transition from the original benign dermoscopic image to the AI-generated malignant counterpart. This results in a brief video, illustrating the subtle progression of skin changes, potentially enhancing patient understanding of skin cancer indicators. Figure 2 displays the transformation for another example.