Cutaneous squamous cell carcinoma ranks as the second most common skin malignancy in Sweden, trailing only basal cell carcinoma. Its incidence is rising sharply, with over 10,000 new cases diagnosed annually across the country. cSCC primarily affects sun-exposed anatomical regions such as the face and neck, where cumulative ultraviolet (UV) radiation inflicts molecular damage over many years, initiating oncogenic mutations in keratinocytes—the predominant cell type in the epidermis.

Dr. Sam Polesie, an associate professor specializing in dermatology and venereology at the University of Gothenburg, led the research team that spearheaded the AI model’s development. “The pathogenesis of squamous cell carcinoma is intimately linked to chronic UV exposure, which induces mutagenic alterations in skin cells. Clinically, these tumors often present on sun-damaged skin characterized by irregular pigmentation, thickened, ulcerated lesions, and compromised elasticity,” explains Polesie. Despite the relative ease of recognizing the disease itself, stratifying tumors according to their biological aggressiveness remains a formidable clinical challenge.

Current medical protocols in Sweden and many other countries generally forgo preoperative punch biopsies in suspected cSCC cases. Instead, patients undergo surgery based solely on clinical suspicion, and the excised tissue undergoes subsequent histopathological examination to confirm diagnosis and guide follow-up care. However, without a biopsy guiding initial treatment, clinicians face significant uncertainty regarding the tumor’s growth potential, complicating surgical planning. Aggressive tumors necessitate prompt, extensive surgical excision with wider margins to minimize recurrence risk, whereas less aggressive lesions may be addressed using more conservative approaches.

Acknowledging these diagnostic hurdles, the research team focused on leveraging machine learning to analyze a robust dataset comprising 1,829 close-up clinical images of pathologically confirmed squamous cell carcinomas. The AI was trained to categorize tumors into three distinct tiers of aggressiveness based on morphological features extracted through advanced image processing algorithms. Subsequent validation utilized an independent test set of 300 images, comparing the AI’s predictive accuracy directly against evaluations performed by a cohort of seven experienced dermatologists.

The study, published in the Journal of the American Academy of Dermatology, revealed that the AI model’s performance in differentiating tumor aggressiveness was statistically indistinguishable from that of the expert clinicians. Intriguingly, the interobserver variability among dermatologists themselves was only moderate, highlighting inherent subjectivity and complexity in human assessments. These findings illustrate that AI can standardize and potentially enhance diagnostic reliability in preoperative settings, offering consistent and swift decision support.

Among the salient clinical indicators identified as correlating strongly with aggressive tumor phenotypes were the presence of ulcerations and flat, non-elevated skin surfaces. Tumors with these characteristics exhibited more than double the likelihood of belonging to higher aggressiveness categories. This insight underscores the importance of integrating specific morphological cues into computational models to refine prognostic accuracy.

Despite mounting enthusiasm for AI applications in dermatology, practical integration into routine clinical workflows has been limited to date. Polesie emphasizes that successful adoption hinges upon targeting AI development toward well-defined clinical problems where enhanced decision-making can tangibly improve patient outcomes. “Our focus has been the preoperative assessment of suspected skin cancers, an area ripe for AI’s capabilities. While our model requires further validation and optimization, its value lies in augmenting—not replacing—clinical expertise,” he asserts.

The AI employed convolutional neural networks (CNNs), a deep learning architecture adept at recognizing intricate spatial patterns within images. By training on a sizeable and diverse image repository gathered between 2015 and 2023 at the Sahlgrenska University Hospital dermatology department, the system learned to discern subtle textural and color variations indicative of tumor biology. This data-driven approach represents a leap beyond traditional diagnostic heuristics, harnessing computational power to harness cellular and tissue-level heterogeneity captured visually.

From a technical perspective, the model’s success rests on sophisticated preprocessing steps—such as normalization, segmentation, and augmentation—that enhance image quality while mitigating noise and variability inherent to clinical photography. The subsequent feature extraction phase employs filters designed to capture edges, gradients, and color contrasts associated with pathological changes. Finally, classification layers map these features onto clinically relevant aggressiveness labels, delivering probabilistic outputs that clinicians can interpret alongside their assessments.

Looking forward, the integration of this AI tool into teledermatology platforms and mobile diagnostic applications could democratize access to expert-level evaluation, particularly in resource-limited settings or remote areas. Rapid, noninvasive tumor characterization facilitated by image-based algorithms might reduce unnecessary surgeries or expedite intervention for high-risk lesions, substantially improving patient care efficiency.

Given the complexity and variability of cSCC presentation, additional studies encompassing larger, multi-center image datasets and diverse patient populations are essential. Moreover, longitudinal analyses linking AI-predicted aggressiveness with actual clinical outcomes will validate the prognostic utility of these computational assessments.

In conclusion, this research from the University of Gothenburg marks a pivotal step toward harnessing artificial intelligence as an adjunct diagnostic tool in dermatological oncology. By matching expert dermatologist performance in gauging squamous cell carcinoma aggressiveness through noninvasive imaging, the AI model demonstrates immense potential to refine surgical planning, optimize resource allocation, and ultimately improve prognosis for thousands of patients afflicted by this widespread cancer. Continued technological refinement paired with clinical validation will pave the way for AI’s meaningful integration into standard dermatological practice.

Subject of Research: People

Article Title: Assessing Differentiation in Cutaneous Squamous Cell Carcinoma: A Machine Learning Approach

News Publication Date: 1-Aug-2025

Web References:
10.1016/j.jdin.2025.07.004

References:
Journal of the American Academy of Dermatology

Image Credits: Photo: Johan Wingborg

Keywords: artificial intelligence, squamous cell carcinoma, skin cancer, dermatology, machine learning, tumor aggressiveness, convolutional neural networks, preoperative assessment, medical imaging, cancer diagnostics

Whole slide imaging has transformed pathology by digitizing glass slides at ultra-high resolutions, often generating images spanning billions of pixels. These gigapixel images provide the detailed morphological information vital for accurate diagnosis but pose substantial challenges for both human interpretation and computational analysis due to their size and complexity. Traditional image analysis methods struggle to process gigapixel WSIs efficiently, often requiring downscaling or patch-based approaches that risk losing critical contextual information. HistoGPT, however, capitalizes on a novel architecture that can ingest these massive images in their entirety, maintaining spatial coherence and enabling comprehensive analysis.

At the core of HistoGPT lies an advanced adaptation of generative pre-trained transformer (GPT) architectures, originally developed for natural language processing tasks. By integrating vision transformer models with generative language models, the researchers have engineered a system that not only interprets visual data from WSIs but also translates these complex imagery patterns into coherent, detailed, and clinically relevant pathology reports. This multimodal learning approach marks a significant leap, transforming image data directly into text with a high degree of fidelity and nuance.

One of the key technical innovations underpinning HistoGPT is its ability to handle hierarchical image representations, enabling it to zoom in and out within the gigapixel WSIs to detect features at multiple scales—ranging from cellular structures to larger tissue architecture. This hierarchical processing mimics the diagnostic approach of human dermatopathologists, who shuttle between high magnification for cellular detail and lower magnification for tissue context. Such a capability ensures that diagnostic reports generated by HistoGPT incorporate microscopic pathological features alongside broader tissue-level abnormalities.

Training HistoGPT required the assembly of a vast and expertly annotated dataset of dermatopathology WSIs paired with corresponding diagnostic reports. The meticulous curation of this dataset was essential not only for teaching the model the complex morphological signatures of diverse dermatological conditions but also for enabling it to learn the language conventions and report structuring used by clinical pathologists. The model’s training regimen involved pre-training on visual and textual data separately before fine-tuning on the integrated multimodal task, a process that substantially enhanced its understanding and fluency in both image interpretation and medical storytelling.

Performance evaluation of HistoGPT demonstrated remarkable results. When benchmarked against human dermatopathologists, the AI system generated reports with substantial concordance in diagnostic terminology, lesion characterization, and treatment recommendations. Importantly, the model achieved this level of performance while operating considerably faster than traditional manual workflows, highlighting its potential to accelerate diagnostic processes in busy clinical environments without compromising quality.

Beyond mere accuracy, HistoGPT exemplifies explainability and transparency, two critical attributes necessary for clinical AI integration. The system is equipped with attention visualization tools that allow users to identify which regions of an image contributed most heavily to specific parts of the generated report. This feature fosters trust among healthcare professionals, ensuring that AI-generated insights can be readily verified and contextualized alongside pathologists’ expertise.

The implications of HistoGPT extend far beyond dermatopathology. As a proof-of-concept for AI-enabled report generation directly from gigapixel WSIs, it lays the groundwork for analogous applications in other pathology subfields, including hematopathology, neuropathology, and oncologic pathology. Each of these disciplines grapples with the dual challenges of large image datasets and complex diagnostic narratives, making HistoGPT’s framework broadly relevant and adaptable.

Moreover, the scalability of this system holds promise for addressing disparities in diagnostic expertise globally. In regions where there is a shortage of highly trained dermatopathologists, AI systems like HistoGPT could act as diagnostic force multipliers, providing high-quality assessments and reports that might otherwise be inaccessible. Such democratization of dermatopathological expertise could lead to earlier diagnoses, improved patient outcomes, and more equitable healthcare delivery worldwide.

However, the integration of AI systems like HistoGPT into routine clinical practice will necessitate stringent validation protocols, regulatory approval, and ongoing surveillance to ensure safety and efficacy. Ethical considerations, including data privacy, informed consent, and the mitigation of algorithmic bias, must be thoroughly addressed before widespread deployment. The authors of this study emphasize collaborative efforts between AI specialists, clinicians, and policymakers to create robust frameworks for responsible AI implementation in healthcare.

Interestingly, HistoGPT’s approach of directly linking raw imaging data to textual reports also presents opportunities for enhancing medical education. By generating detailed and annotated reports from complex WSIs, such systems could serve as interactive teaching tools to train both pathology residents and practicing clinicians, exposing them to diverse case presentations and diagnostic reasoning paths in a highly accessible format.

The research team envisions future iterations of HistoGPT incorporating multimodal data beyond histological images, potentially integrating genomic, proteomic, and clinical metadata to create even richer diagnostic narratives. This comprehensive approach aligns with the growing trend toward precision medicine, where multi-dimensional data synthesis informs tailored therapeutic strategies and prognostic assessments.

As the boundaries of AI and medical imaging continue to blur, HistoGPT represents a compelling example of how transformer-based architectures and deep learning can bridge the gap between visual data comprehension and natural language generation in clinical workflows. Its success heralds a new era where AI not only supports but actively participates in the complex cognitive tasks of medical diagnosis and reporting.

In conclusion, the arrival of HistoGPT is a landmark moment in computational pathology. By effectively translating gigapixel dermatopathology WSIs into structured, accurate, and clinically meaningful reports, it promises to transform diagnostic pathology from a largely manual, labor-intensive endeavor into a streamlined, AI-augmented discipline. As ongoing research refines and validates this technology, patients and clinicians alike stand to benefit from faster, more precise dermatological diagnoses, and a future where AI becomes an indispensable partner in personalized medicine.

Subject of Research: Automated generation of dermatopathology diagnostic reports from gigapixel whole slide images using a transformer-based AI system.

Article Title: Generating dermatopathology reports from gigapixel whole slide images with HistoGPT.

Article References:
Tran, M., Schmidle, P., Guo, R.R. et al. Generating dermatopathology reports from gigapixel whole slide images with HistoGPT. Nat Commun 16, 4886 (2025). https://doi.org/10.1038/s41467-025-60014-x

Image Credits: AI Generated