Advances in dermatopathology
Dermatopathologists discuss the latest trends and developments in dermpath.
By Allison Evans, Assistant Managing Editor, November 1, 2023
The field of dermatopathology is experiencing much growth and change. From how to use PRAME to aid in determining whether a lesion may be malignant, to digitizing glass slides and being able to use validated data sets for artificial intelligence programs, these changes are just starting to make ripples for dermatopathologists and dermatologists alike.
Dermatopathologists discuss recent and upcoming advances in dermatopathology and what dermatopathologists and dermatologists should know.
Short on time?
Key takeaways from this article:
In recent years, PRAME has come forth as a potential antigen to help distinguish melanomas (PRAME+) from benign melanocytic neoplasms (PRAME−). With the availability of PRAME immunohistochemistry (IHC), pathologists have been very busy in the last several years evaluating its potential diagnostic role in melanoma.
PRAME is not a substitute for routine histology and other melanocytic markers that are more commonly used such as Melan-A, HMB-45, and SOX10. It’s an adjunct to the histology and traditional melanocytic markers.
Although the technology for whole-slide imaging is increasingly being used in teaching and some consulting scenarios, digital pathology must overcome some significant obstacles before making the leap to assisting in the diagnosis of patients in the office.
While there are some barriers to adopting WSI, such as cost, there are significant benefits, including reduced turnaround time, improved convenience, and expanded access to experts.
In dermatopathology, AI can recognize routine diagnoses, supporting intraoperative consultation during Mohs surgery, and classifying melanocytic tumors.
Studies so far show AI outperforms dermatologists in the diagnosis of skin cancers and combining the performances of both results in even better performance without a decrease in specificity.
Limitations of AI include lack of standardized, validated images, mostly binary diagnostic classification system, and systematic errors in algorithms.
Melanocytic neoplasms are known to be diagnostically challenging. PRAME (preferentially expressed antigen in melanoma) is the latest immunohistochemical (IHC) stain being used by dermatopathologists. PRAME, in contrast to other melanocytic markers, is differentially expressed in melanocytic nevi and melanomas. In recent years, PRAME has come forth as a potential antigen to help distinguish melanomas (PRAME+) from benign melanocytic neoplasms (PRAME−). With the availability of PRAME immunohistochemistry, pathologists have been very busy in the last several years evaluating its potential diagnostic role in melanoma.
“Overexpression of PRAME is observed in melanomas and a variety of other cancers including breast, lung, kidney, ovary, and leukemias,” said Jason B. Lee, MD, FAAD, director of the Jefferson Dermatopathology Center, Dermatology Fellowship, and Jefferson Pigmented Lesion Clinic at Thomas Jefferson University. PRAME is now available for dermatopathologists doing an immunohistochemical stain; simply apply the stain and see the presence or absence of PRAME. PRAME ≤25% (0/1+) favors nevus, PRAME 26%-75% (2/3+) is noncontributory, and PRAME >75% (4+) favors melanoma.
“PRAME is not a substitute for your routine histology and other melanocytic markers that are more commonly used such as Melan-A, HMB-45, and SOX10,” said Alina G. Bridges, DO, FAAD, director of dermatopathology and associate professor in the department of dermatology and anatomic pathology at Hofstra/Northwell. “It’s an adjunct to the histology and traditional melanocytic markers.”
“There are pitfalls dermatopathologists need to know about, including how to interpret the stain correctly. The most helpful usage of it is if you have a borderline lesion and you are trying to use the stain to help you distinguish severely dysplastic or atypical nevus from melanoma,” she added.
“PRAME can also help if you see a nevoid melanoma that on low power has an architecture that resembles a nevus. But when you look closer, you start seeing that it has mitoses and atypia,” Dr. Bridges said. “It can also help if you’re looking at a sentinel lymph node and trying to figure out if it’s a nevus in the lymph node versus metastasis. But it will only work if the primary tumor is PRAME positive.”
“There are pitfalls dermatopathologists need to know about, including how to interpret the stain correctly. The most helpful usage of it is if you have a borderline lesion and you are trying to use the stain to help you distinguish severely dysplastic or atypical nevus from melanoma.”
Another utility, continued Dr. Bridges, is “a lot of times when you’re trying to look at margins, particularly for melanoma in situ on chronic sun-damaged skin, it’s hard to sometimes clear margins because there’s increased melanocytes on sun-damaged skin and you don’t know if they’re benign or malignant, and so using the stain may help you recognize that it’s just melanocytic hyperplasia from sun-damaged skin and not residual melanoma at the margin.”
“Now if PRAME is present, does it always mean it’s melanoma? The answer is no. If it’s absent, is it not a melanoma? This is not true either,” Dr. Lee said. “It has some sensitivity and specificity issues.” About 14% of benign nevi will be PRAME positive and 15% of solar lentigines will also be PRAME positive, Dr. Bridges added. “PRAME has the lowest positivity for desmoplastic melanoma at 35%.”
PRAME has been found to be consistently present in metastatic melanomas, said Dr. Lee. “For thin melanomas, we’re trying to figure out how often it’s expressed. Down the road, we need more objective studies conducted independently verifying their data.”
“In one of the largest iterations of the PRAME IHC studies evaluating over 400 melanocytic lesions, the sensitivity in detecting melanoma ranged from 83% to 94% depending on the subtype of melanoma. Slightly lower sensitivity was observed in difficult-to-diagnose melanocytic lesions,” explained Dr. Lee.
“With encouraging results thus far, pathologists are rapidly incorporating this readily available, cost-effective, and fast-turnaround test into their daily practice,” Dr. Lee said. “Straightforward nevi and melanomas should not require PRAME IHC — adding only unnecessary cost if indiscriminately utilized. The true value of PRAME IHC should be judged on its performance on the rarer difficult-to-diagnose lesions where a pathologist needs an additional confirmatory test.”
“The current research only supports the value of PRAME as an ancillary tool to evaluate challenging melanocytic lesions. It’s to be used in combination with expert morphologic examination, other available testing, and clinical context,” Dr. Bridges said.
“Ambiguous melanocytic lesions remain a real challenge in dermatopathology, and IHCs like PRAME will need additional studies to determine optimal utilization,” Dr. Lee noted.
Virtual microscopy, also known as whole slide imaging (WSI), is the process whereby, with scanning machinery, you’re able to take a whole mount of a glass slide and get digital imagery of that entire specimen and then, with a computer program, convert that into an image that can be placed online.
Over the years, significant gains in technology have led to the adoption of innovative digital imaging solutions in pathology, which continue to evolve at a rapid pace. Greater computer processing power, data transfer speeds, advances in software, and cloud storage solutions have enabled the use of digital images for a wide variety of purposes in pathology, including diagnostic, educational, and research. “Although the technology is increasingly being used in teaching and some consulting scenarios, it must still overcome some significant obstacles before making the leap to assisting in the diagnosis of patients in the office,” Dr. Lee said.
Digitizing slides takes an enormous amount of data space, Dr. Lee said. “Just one slide will take about a gigabyte or two, depending on the size of the tissue. A small lab will have about a minimum of 30,000 slides where larger laboratories can have 300,000 specimens, which equates to a lot of digital storage space.”
Dr. Lee tried to convert his laboratory to digital several years ago. However, the cost was prohibitive. “The entire IT infrastructure needed to be revamped, so a lot of resources need to go into it. Cost is a huge issue.”
Although there are companies selling systems of hardware and software for digital pathology, it’s several millions of dollars to purchase and use the products — and then you must buy into their operational ecosystem, he added.
But as time goes on, the technology gets cheaper, the digital storage space gets cheaper, so eventually it will be done, Dr. Lee noted. “It’s just a matter of time, but it’ll be slow. As of now, I only know of a few institutions that have converted to entirely digital pathology.”
Another concern that arises from the use of WSI is the accuracy of interpretation from the whole slide images and the effect on workflow. While most of the literature show that diagnoses can indeed be rendered by WSI, an error rate of approximately 1-5% for WSI does exist, although there is also a low baseline discrepancy rate for glass-to-glass slide review (doi: https://doi.org/10.2147/PLMI.S59826).
There are a lot of data to show that the imaging on a digital slide read is the equivalent to that of a microscope, according to Thomas Olsen, MD, FAAD, lab director of the Dermatopathology Laboratory of Central States in Dayton, Ohio, in a 2016 DermWorld article. Dr. Olsen, whose laboratory has a considerable investment in digital technology, reported that he could make a digital diagnosis approximately 95% of the time. For the remaining 5%, he would review the glass slide under a microscope to identify with more certainty different nuances of the case.
The time required to review digital slides may be a limiting factor in the widespread adoption of WSI. Since smaller specimens (e.g., biopsies) occupy less area on a slide, they are generally easier to scan. It takes less time to digitize small areas of interest than an entire glass slide. Also, with large specimens, tissue sections on the slide may extend beyond the coverslip, making it hard to focus. It is also important to be aware that very small pieces of tissue may not always be captured with WSI, especially with faint staining (https://doi.org/10.2147/PLMI.S59826).
“The technology is getting better and faster,” said Dirk Elston, MD, FAAD, chair of the dermatology and dermatologic surgery department at the Medical University of South Carolina. “But for most of us it’s still nowhere near as fast and cheap as using a glass slide,” he said. “It’s not that I don’t think digital pathology is wonderful. It’s just that cost will keep it from replacing H&E slides any time soon.”
“While there are still many barriers to the widespread adoption of digital pathology, including lack of reimbursement, it’s important to recognize where dermatopathology is heading and bridge its past with its future,” said Warren Heymann, MD, FAAD, in a 2016 issue of DermWorld Insights and Inquiries. “Does your radiologist hold up a film to an X-ray box? It won’t be long before the microscope is a relic.”
Putting care into context
New dermatopathology appropriate use criteria from ASDP help guide patient care. Read more.
The use of digital slides reduces the wear and tear on glass slides as well as the physical obstacles of transporting them, said Dr. Bridges. “Because the speed of processing and memory has improved so dramatically in recent years, digital slides can be stored quickly and easily without taking up any physical space. By uploading them at high resolution, they are preserved in perpetuity. With this technology, sharing interesting cases is no longer limited to the individuals in possession of the glass slide,” she added.
“When you digitize slides, they’re readily available, as opposed to glass slides which have to be stored in a physical location,” Dr. Lee said. “They can be easily shared within your institution or anywhere in the world, which aids with consultations.”
“Digitizing slides improves teaching and research. There are more advantages and opportunities in digital pathology than disadvantages. You can improve efficiency, reduce turnaround times, and have better coordination of care between yourself and other physicians, health care workers, and patients,” Dr. Bridges said.
“There is improved ability to share cases and do consultations from places where patients might not have access to a dermatopathologist — more rural, remote areas,” said Dr. Bridges. “There are also global opportunities not just for consults, sharing opinions, and telepathology, but also research opportunities. That leads to AI because now you have data and you can build AI tools that will help with diagnosis, prognosis, and treatment, and eventually predict responses to treatment.”
Another advantage is that the slides remain intact. “One of the issues I have is that I collect slides for teaching but slides break or they fade over time. I now have thousands of digitized slides I’ve been collecting for teaching purposes and publications,” Dr. Lee said.
For Dr. Bridges, digital pathology improves her work-life balance because she can take it anywhere. “You can do this remotely. During the pandemic, CMS loosened some of the regulations and now CMS is allowing the remote process of signing out cases to continue.”
Experts say that as more image-analysis algorithms and computer-assisted diagnosis tools get developed and validated for clinical use, they may empower pathologists to become more efficient, precise, and reproducible at quantifying prognostic biomarkers. While there are some barriers to adopting WSI, experts say there are significant benefits: reducing turnaround time, improving convenience, and expanding access to experts.
Batch effects and hidden variables in AI
“We want to use diverse data sets to train algorithms, but this introduces the tremendous potential for batch effects and hidden variables,” said Dr. Motaparthi. “Hidden variables are present in heterogeneous data sets, which result in batch effects. A hidden variable is something that you see due to technological artifacts, usually, but can also be biologic changes. Patient age, what laboratory the specimen came from, the H&E processing protocol, type of scanner used — those are all hidden variables.”
“If you look at the same slide from different scanners, they can look very different to your eye,” Dr. Motaparthi explained. “They also look different to the algorithm, too. The algorithm starts to learn those hidden variables rather than learning the target variable. You might say that a lesion is melanoma because a hotspot on the slide shows pagetosis or cytologic atypia. But the algorithm learns that a certain H&E protocol, or another hidden variable, is a reason for why it’s melanoma. It will use that variable next time to determine whether a slide is melanoma, and it won’t be because of the reason that you want but because of what it learned from the hidden variable.”
“We want tons of opinions from different people based on lots of slides from different laboratories,” Dr. Motaparthi said. “This seems like the optimal way to do it, but when you’re training an algorithm, it leads to batch effects. To get past this, we’ll need to standardize processing, standardize the way we digitize slides, and use really large samples.”
AI in dermatopathology
Combining WSI with image analysis tools allows users to leverage technology to perform tasks that were previously too cumbersome or even impossible for humans to undertake manually. Artificial intelligence (AI) refers to the ability of computer systems to perform tasks that traditionally require human input. AI, including deep learning methods that leverage neural network-based algorithms, could hold significant promise for dermatopathology.
“Progress has been limited for several decades until recently with advances in machine learning, a subset of AI in which the computer algorithms make predictions without being explicitly programmed,” Dr. Lee said. “Deep learning entails the creation and layering of multiple artificial neural networks that, when given data to review, allows for self-training and can improve the system’s accuracy.”
When images are used as inputs, they are broken down into pixels that are individually analyzed by the deep neural network, similar to the technology underlying many facial recognition applications, wrote dermatopathologist Kiran Motaparthi, MD, FAAD, and colleagues in an article published in the Journal of Cutaneous Pathology. Given the ability to extract complex image features and make predictions without human intervention, convolutional neural networks are ideal for dermatopathology.
“In dermatopathology, AI can recognize routine diagnoses — supporting intraoperative consultation during Mohs surgery, and even classifying melanocytic tumors,” said Dr. Motaparthi, director of dermatopathology and clinical professor of dermatology at the University of Florida College of Medicine. “The most important role of AI is to distinguish malignant and benign pigmented lesions as the approach to treatment varies significantly based on the classification” (https://doi.org/10.1111/jocd.15565).
“AI has already proven useful in tedious tasks that are prone to variability, such as counting mitoses and quantifying HER2 expression in breast cancer. Naturally, applications of AI have now extended to dermatopathology,” Dr. Motaparthi said, “in which it may help with surgical margin assessment, identifying lymphovascular invasion, diagnosis of onychomycosis, and even drafting reports. Having algorithms for these kinds of tasks could make us more available for challenging, high-level tasks. Deep learning algorithms have already been trained to recognize basal cell carcinoma, dermal nevus, and seborrheic keratosis at an accuracy over 99%, suggesting the utility of AI for diagnosis of very common entities.”
There is a wealth of data demonstrating a degree of promise for AI in the diagnosis of skin conditions. Xie et al utilized a set of over 2,000 histopathological images to assess the accuracy of two deep learning architectures in distinguishing melanoma from nevi, finding high overall accuracy, including 92% sensitivity and 94% specificity. Similarly, Cruz-Roa et al used a set of 1,417 histopathological images to create a deep learning architecture that distinguishes basal cell carcinoma from normal tissue with an accuracy of 91.4%.
“In dermatology, advances in digital cameras and dermoscopy devices have contributed to a significant increase in the amount of digital data available. Deep learning algorithms can now distinguish dermoscopic images of nevus and melanoma with accuracy comparable to or exceeding that of junior and senior dermatologists,” Dr. Motaparthi added. A study published in the European Journal of Cancer compared the performance of a deep-learning algorithm trained by open-source images exclusively compared to many dermatologists covering all levels within the clinical hierarchy. The mean sensitivity and specificity achieved by the dermatologists with dermoscopic images was 74% and 60%, respectively. Among the dermatologists, the chief physicians showed the highest mean specificity of 69.2% at a mean sensitivity of 73.3%. With the same high specificity of 69.2%, the convolutional neural network (CNN) had a mean sensitivity of 84.5%. The CNN outperformed 136 of the 157 dermatologists at all the different levels of experience.
“Currently, deep learning is being evaluated for its potential to produce clinically meaningful classification of melanocytic tumors. If successful, AI could reduce clinically impactful interrater discordance in the interpretation of melanocytic neoplasms,” said Dr. Motaparthi. “Diagnosis of melanocytic tumors is subject to high interrater variability. A deep learning algorithm has been trained to distinguish nevi from melanomas based on cropped diagnostic portions of whole slide images. Surprisingly, the discordance rate between the algorithm and the ground truth diagnosis was less than the discordance between a board-certified dermatopathologist in earlier studies.”
“AI outperforms dermatologists in the diagnosis of skin cancers and combining the performances of them results in even better performance without a decrease in specificity,” said Dr. Lee. “These studies show that AI has potential to improve the diagnosis and management of skin cancers.” Dr. Lee notes, however, that these results are based on artificially controlled study settings and have not been tested in a real-world setting.
“Deep learning algorithms can also differentiate conventional nevi from Spitz nevi and normal skin from melanoma with an accuracy of 92% and 96%, respectively,” Dr. Motaparthi said. Reliant only upon a single H&E-stained slide, a deep learning-based biomarker accurately correlates with higher risk of metastasis and poorer disease-specific survival in early-stage melanoma (doi: 10.1158/1078-0432.CCR-19-1495).
While most findings are based on algorithms capable of binary classification, recently, three sequential algorithms were capable of placing a single slide into one of four categories with an accuracy rate of up to 90%. “While these broad classifications (basal cell carcinoma-like, squamous cell carcinoma-like, melanocytic, and other) are not practically useful diagnoses, this finding suggests AI may eventually produce a tool capable of triage in dermatopathology,” Dr. Motaparthi noted. “AI-based triage could translate to reduced turnaround time, a prioritized workload in which we review our unclassifiable cases first, or a workflow that allows us to identify our most challenging cases for ancillary testing or expert consultation.”
Augmented artificial intelligence
Read the Academy’s Position Statement on Augmented Artificial Intelligence.
“An algorithm that you train and validate and then test is only as good as the ground truth you give it,” said Dr. Motaparthi. “There’s a large potential for very frustrating systematic error when we have conflicting ground truths or poorly curated ground truths. For instance, there are algorithms that combine pathologic and clinical data, but some of the clinical images could be diagnoses that were not biopsy verified. You have to train using the same ground truth that you also test; otherwise, you get very frustrating results.”
“Dermpath does not have a large, validated pool of images where anyone can access the data and perform analyses. Everyone is publishing their own datasets and claiming that it works, so it’s not generalizable,” Dr. Lee said. “This is why we don’t yet have an FDA-cleared product for dermatopathology,” he added, “unlike radiology which has a very large pool of validated images and an FDA-cleared product. Artificial intelligence requires lots of data. Data is like the new ‘oil’ for the future; everyone wants it.”
Also, the degree of image sharing among sources is poor and the quality of images is not uniform. For effective functioning of deep learning algorithms, there is a need for substantial quantity of diverse and high-quality data to improve diagnostic efficacy (https://doi.org/10.1111/jocd.15565).
The sole reliability of AI in dermatopathology practice remains problematic. CNNs have been geared to recognize limited types of lesions, such as nevi, seborrheic keratoses, and basal cell carcinoma. Many lesions biopsied in the day-to-day clinical setting fall outside of these categories. Furthermore, dermatopathological diagnoses do not always correlate clinically. In addition, histopathologic diagnoses closely mimic one another in clinical practice and require nuanced analysis (https://doi.org/10.3390/dermatopathology10010014).
“Currently, deep learning is primarily capable of binary diagnostic classification in dermatopathology. Before AI can be employed in practice, a deep learning system that produces multiclass diagnosis is needed,” wrote Drs. Lee, Motaparthi, and colleagues in an article published in the Journal of Cutaneous Pathology. “This holistic model should be tested on images from multiple labs with variation in tissue processing, a wide spectrum of pathology, and artifacts that represent a typical diagnostic workload in dermatopathology.”
Dr. Lee believes AI in dermatopathology can help with repetitive tasks. “AI can triage a lot of these routine diagnoses like cysts and skin tags, and it can prioritize diagnoses like melanoma and skin cancers. It can also aid with the binary system of diagnosis. For example: Is it cancer or not? It could be used as an adjunctive test for melanoma.”
“Right now, if I were having difficulty and wanted to test, I might send the specimen out for gene expression profiling and they would give me a score that corresponds to benign, malignant, or unsure,” Dr. Lee said. “Instead of using this testing, which is costly and takes a lot of time, we could just run an AI program that would hopefully be just as good or better. AI may bring cost effectiveness to the laboratory once it’s all implemented.”
As the technology becomes more refined, it may be used to perform preliminary reads, similar to cardiologists using an electrocardiogram. This approach may prove helpful in practices with a high volume of slides and paucity of trained dermatopathologists, helping to ensure patients with malignant lesions are diagnosed in a timely manner (https://doi.org/10.3390/dermatopathology10010014).
“Deep learning algorithms are dependent on dermatopathologists for the ground truth diagnoses that provide training and validation. We should begin to consider AI as an assistant — to improve efficiency, reduce tedium, and perform triage — not as a replacement,” Dr. Motaparthi said.
“People typically underestimate the abilities of AI,” Dr. Lee said. “It will eventually come to us, and once it gets established, the change is going to come very quickly. I recommend those in the field of dermatology pay attention to what’s going on in the sphere of AI.”
Dr. Elston quoted colleague Victor Prieto, MD, PhD: “‘AI is the cheapest fellow you will ever hire; never sleeps; always on time; sometimes lies.’ AI is a hard worker, but you have to double check the work.”