Abstract
Representation of diverse skin tones in medical literature is critical for equitable healthcare, yet manual auditing of educational materials remains a significant bottleneck. This paper presents STAR-ED, an automated machine learning framework designed to detect and classify skin tones within medical textbooks. By processing raw PDFs through a pipeline of document ingestion, image selection, and deep learning-based classification, STAR-ED provides a scalable method for diversity assessment. Our analysis of four gold-standard medical textbooks reveals that only 10.5% of images represent dark skin tones (Fitzpatrick scales V-VI), underscoring a systemic bias that may contribute to diagnostic disparities in clinical practice.
Introduction
Medical education faces a persistent representation challenge. Historically, dermatology and primary care textbooks have predominantly featured light skin tones (Fitzpatrick scales I-III). Because clinical presentations of various conditions—ranging from inflammatory rashes to malignancies—can vary significantly based on skin pigmentation, this lack of diversity leaves medical professionals undertrained in recognizing diseases on darker skin.
Quantifying this bias has traditionally required domain experts to manually review thousands of pages, a process that is time-consuming, subjective, and difficult to scale. The STAR-ED framework was developed to address this by providing an objective, automated, and scalable tool for publishers and educators to evaluate the inclusivity of their materials.
Methodology
The STAR-ED framework employs a multi-stage pipeline to transform unstructured PDF documents into comprehensive diversity reports.
- Document Ingestion: The system utilizes the Corpus Conversion Service (CCS) to parse PDFs, extracting text, tables, and images into a structured JSON format.
- Skin Image Selection: To filter out non-clinical images such as diagrams or pathology slides, an XGBoost classifier was trained using Histogram of Oriented Gradients (HOG) and color statistics in the CIE LAB color space.
- Skin Pixel Segmentation: Intensity-based masking is applied to isolate skin pixels, minimizing interference from background elements like clothing or medical equipment.
- Skin Tone Estimation: The core classification is performed by a ResNet-18 model, pretrained on ImageNet and fine-tuned on the Fitzpatrick17k dataset. The model categorizes skin tones into two groups: Light (Fitzpatrick I–IV) and Dark (Fitzpatrick V–VI).
This deep learning approach proved more robust than traditional metrics like the Individual Typology Angle (ITA), which often struggles with variations in lighting and image quality.
Datasets and Validation
The framework was developed and validated using several datasets:
- Fitzpatrick17k: 16,577 clinical images used for training the skin tone estimator.
- DermEducation: 2,708 images used to train the skin-vs-non-skin classifier.
- Medical Textbooks: Four major textbooks, including Rook’s Textbook of Dermatology and Fitzpatrick’s Dermatology in General Medicine, served as the primary test bench.
Validation against expert-provided manual labels confirmed that the automated framework serves as a reliable proxy for human assessment.
Results
The STAR-ED framework demonstrated high performance across its primary tasks:
- Skin Detection: The XGBoost classifier achieved an AUROC of 0.96 and an F1 score of 0.90.
- Tone Classification: The fine-tuned ResNet-18 model achieved an AUROC of 0.87 and an F1 score of 0.91.
The most significant finding was the quantitative confirmation of underrepresentation: across all analyzed textbooks, the proportion of dark skin tones (FST V-VI) never exceeded 10.5%. This provides a clear, data-driven benchmark for the current state of medical publishing.
Discussion and Limitations
STAR-ED enables "continuous integration" for diversity in publishing, reducing the time required for a full textbook audit from over 100 person-hours to just minutes. However, several limitations remain. The current binary classification (Light vs. Dark) lacks the granularity of the full 6-point Fitzpatrick or Monk scales. Additionally, diseased skin can sometimes interfere with accurate segmentation, and the model remains sensitive to the lighting conditions of the original photography. Future work will focus on integrating contextual text analysis and improving segmentation for lesional skin.
Summary updated