The lack of darker skin images in dermatology educational materials is a significant issue and may translate to disparities in care and hence outcome. To address this issue, we propose a framework to automatically assess bias in skin tone representation in academic documents of dermatology. Given a document, we apply content parsing to extract text, images, and table cells in a structured format. We then select skin images and segment non-disease regions using Mask R-CNN. Individual Typology Angle (ITA) values are computed from non-disease regions and mapped to Fitzpatrick skin indices. The proposed framework is validated with three dermatology textbooks and compared against manually annotated baselines by dermatology experts. Results show encouraging performance in estimating skin tones and discover limited representation of darker skins, i.e., only 10.7%, across these documents. We envision this technology as a tool for dermatology educators to quickly assess their materials.