AI in SK Histopathological Diagnosis (AI\SK)

Seborrheic keratosis (SK) is one of the most common benign epidermal tumors. Treatment is generally unnecessary, although lesions may be removed because of irritation, pruritus, or cosmetic concerns. SK has several clinical and histological subtypes, including common seborrheic keratosis (CSK), which is more prevalent among Caucasians, and dermatosis papulosa nigra (DPN), which is more common in individuals with Fitzpatrick skin phototypes III and above. Its development is associated mainly with age and genetic predisposition, with possible contribution from ultraviolet radiation. Lesions may occur almost anywhere except the palms and soles, with the face and upper trunk being common sites. Diagnosis is usually clinical but may be supported by dermoscopy or histopathology.

Histologically, SK represents an intraepidermal proliferation of squamous or basaloid cells. The characteristic findings include acanthosis, papillomatosis, hyperkeratosis, keratin cysts, and keratin pseudocysts. Cellular atypia is generally absent, while the amount of melanin and melanocytes varies according to the degree of pigmentation. SK shows considerable clinical and histological variability, which can sometimes make differentiation from other lesions, such as keratoacanthoma and clear cell acanthoma, challenging.

Recent advances in digital pathology and artificial intelligence (AI) have created new opportunities for diagnostic support in dermatopathology. AI aims to mimic aspects of human intelligence, while machine learning enables computers to identify patterns from data. Deep learning, neural networks, and convolutional neural networks (CNNs) are particularly important in image analysis. AI in dermatopathology has evolved from early text-based systems such as TEGUMENT, introduced in 1987, to modern systems capable of directly analyzing digital pathology images. Although widespread clinical implementation is still developing, AI has potential applications in diagnosis, triage, education, and research and requires collaboration between dermatopathologists, pathologists, clinicians, engineers, and data scientists.

Study Methodology

H&E-stained glass slides will be collected from the Al-Hussein dermatopathology archive, covering the period from 2010 to 2019. A panel of certified dermatopathologists will evaluate the slides. Their diagnoses will serve as the gold standard against which the AI results will be compared.

The slides will be digitized using a Leica Aperio GT450 scanner at 40× magnification. Before scanning, slides will be cleaned and assessed for adequate staining, proper coverslipping, and absence of artifacts such as air bubbles, tissue folds, or debris. The scanner can accommodate 15 racks with 30 slides per rack, for a total of 450 slides per run, with approximately 1.5-2 minutes required to scan each slide.

Whole Slide Images (WSIs) will be generated and stored in SVS format, with individual file sizes typically ranging from approximately 500 MB to 2 GB. The digital slides will then undergo quality control using Aperio ImageScope to ensure adequate focus, clarity, and complete tissue capture. Suboptimal slides will be rescanned. The images will subsequently be transferred to an external hard drive for storage and backup before being uploaded to the HistoGPT cloud platform for AI analysis.

HistoGPT Analysis

The WSIs will be analyzed using HistoGPT, a Vision-Language Model for Digital Pathology. Because WSIs are extremely large, they are first divided into smaller image patches or tiles, typically 256×256 or 512×512 pixels. Tissue detection is then performed to exclude background areas and focus the analysis on the tissue.

HistoGPT uses a Vision Transformer (ViT) and a Hierarchical Vision Transformer (HiViT) to extract histological features at multiple scales. This allows the model to identify both microscopic features, such as cellular atypia and mitotic figures, and larger architectural patterns, such as symmetry, circumscription, and infiltrative growth.

Through attention mechanisms, the model can also establish relationships between spatially distant areas of the slide and recognize the importance of tissue organization and anatomical location. Its cross-modal alignment allows visual histological features to be associated with terminology used in pathology reports, enabling recognition of concepts such as keratin pearls, clefting, and solar elastosis.

Finally, the language component of HistoGPT generates a structured pathology report, potentially including a microscopic description, differential diagnosis, and final diagnosis. The AI-generated reports will be compared with the reports and diagnoses produced by the dermatopathology experts, and the time required for the diagnostic process will also be assessed.

Statistical Analysis

The collected data will be analyzed using SPSS version 26.0 or R programming. A p-value < 0.05 will be considered statistically significant. Overall, the study aims to evaluate the diagnostic performance and time efficiency of HistoGPT in dermatopathology by comparing its findings with expert dermatopathologists serving as the gold standard.