Enhancing Diagnosis through AI-driven Analysis of Reflectance Confocal Microscopy

TL;DR

This work addresses improving dermatological diagnosis from Reflectance Confocal Microscopy (RCM) images by combining AI with expert input. It employs a patch-based pipeline using self-supervised learning (DINO) to train a Vision Transformer on $256\times256$ patches with stride $64$ and boundary mirroring, producing $768$-dimensional feature embeddings that are clustered with $k$-means, with the optimal cluster count determined by silhouette scores. Expert dermatologists interpret cluster semantics, linking them to clinical concepts such as regular/irregular epidermis, atypical cells, and artifacts, and visualizing them as cluster maps for potential biopsy guidance. The approach demonstrates a promising, unsupervised pathway to segment and assess severity in RCM images, offering a foundation for decision support and education while highlighting the need for larger datasets and comparative analyses. This work thus advances non-invasive imaging analytics toward practical clinical impact in skin cancer diagnostics.

Abstract

Reflectance Confocal Microscopy (RCM) is a non-invasive imaging technique used in biomedical research and clinical dermatology. It provides virtual high-resolution images of the skin and superficial tissues, reducing the need for physical biopsies. RCM employs a laser light source to illuminate the tissue, capturing the reflected light to generate detailed images of microscopic structures at various depths. Recent studies explored AI and machine learning, particularly CNNs, for analyzing RCM images. Our study proposes a segmentation strategy based on textural features to identify clinically significant regions, empowering dermatologists in effective image interpretation and boosting diagnostic confidence. This approach promises to advance dermatological diagnosis and treatment.

Figures (5)

• Figure 1: Patch generation from RCM images. (a) applying mirroring to the edges of image and (b) making image patches of 256x256 pixels.
• Figure 2: ViT small feature extraction model trained using the image patches from RCM images with the DINO algorithm.
• Figure 3: The silhouette scores of the clusters obtained from the image patch features using the $k$-means clustering algorithm are displayed. The x-axis represents the number of clusters $k$, while the y-axis denotes the silhouette score corresponding to each $k$.
• Figure 4: Example of patch images clustered by their image features determined by the ViT model trained by the DINO algorithm, associated with (a) cluster 0, (b) cluster 1, (c) cluster 2, and up to (r) cluster 17.
• Figure 5: Sample RCM images are superimposed with the cluster map. (a) and (b) biopsy not recommended and (c) and (d) biopsy recommended cases.