Go to AAD Home
Donate For Public and Patients Store Search

Go to AAD Home

Deep learning in dermatopathology

DII small banner

By Kiran Motaparthi, MD
June 30, 2021
Vol. 3, No. 26

Kiran Motaparthi
Deep learning algorithms — deep neural networks that model the connections neurons make and analyze images in successive layers to ultimately produce a likelihood of diagnosis — can now distinguish dermoscopic images of nevus and melanoma with greater accuracy than junior and senior dermatologists. (1) Artificial intelligence (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.

So far, the majority of studies of AI in dermatopathology have been based on binary classification: is a diagnosis or a feature present or absent? 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. (2) AI can also support intraoperative consultation during Mohs micrographic surgery. A recent study demonstrated a sensitivity of 100% and specificity of 94% for the detection of basal cell carcinoma in Mohs slides, an accuracy comparable to that of Mohs surgeons. (3)

Diagnosis of melanocytic tumors is subject to high interrater variability: can AI improve our consistency? 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. (4) Deep learning algorithms can also differentiate conventional nevi from Spitz nevi and normal skin from melanoma with an accuracy of 92% and 96%, respectively. (5, 6) 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. (7)

Of note, all of these findings are based on individual algorithms capable of binary classification. Recently, three sequential algorithms were capable of placing a single slide into one of four categories, with an accuracy up to 98%. While these broad classifications (basal cell carcinoma-like, squamous cell carcinoma-like, melanocytic) are not practically useful diagnoses, this finding suggests AI may eventually produce a tool capable of triage in dermatopathology. 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. (8) 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.

Illustration for DWII Deep learning in dermatopathology

Image from JAAD.

Common fears about AI include a lack of understanding about deep learning, the potential liability for machine errors, and the fear of being replaced by machines. (9) Staying abreast of advances in this emerging field — and contributing to the development of algorithms — can reduce fears. Deep-learning supported diagnosis has to be verified before it is actionable, and ground truth diagnoses that train algorithms are based on concordance or consensus between dermatopathologists. In a study of patients with melanoma, the vast majority of patients would support the use of AI in clinical decision-making and would even make their data available to develop better deep learning algorithms. However, most patients would only favor the use of AI to assist dermatologists, rather than to make independent clinical decisions. (10) Given the significant nuance in diagnosis, and the wide spectrum of diagnoses in dermatopathology, it is unlikely that deep learning algorithms will substitute for experienced dermatopathologists.

Point to Remember: In dermatopathology, AI is capable of recognizing routine diagnoses, supporting intraoperative consultation during Mohs surgery, and even classifying melanocytic tumors. However, this diagnostic capability is currently broad; additionally, 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.

Our expert's viewpoint

Jason B. Lee, MD
Clinical Vice Chair
Director, Jefferson Dermatopathology Center
Director, Residency and Dermatopathology Fellowship
Director, Jefferson Pigmented Lesion Clinic
Jefferson University Hospitals

AI is in its early stage of development in pathology and dermatopathology. As mentioned, narrow-based tasks that are repetitive and mundane, such as quantification pathology, are already augmented by AI in real-world practice. AI-augmented binary classification of melanoma has the potential to increase diagnostic precision and accuracy. The performance of AI, however, will depend on the validity of the training datasets utilized. The concordance rate among pathologists is poor for atypical melanocytic lesions, melanoma in situ, or early melanoma — a vast majority of the melanomas diagnosed today. Thus, the training datasets need to be concordant. While AI tools proficient in classifying a narrow scope of diagnoses may be a reality in the near future, the development of an AI tool that can classify a myriad of diagnoses in dermatopathology and think like a dermatopathologist will not be a reality anytime soon. The current barrier to AI application in pathology is the lack of digital data. While radiology has embraced the digitalization of the specialty, pathology has been slow to embrace the conversion. High cost and disruptive changes in workflow are the major factors that prevent laboratories from converting to a digital workflow; however, the COVID-19 pandemic has provided a renewed impetus to do so. Once pathology converts, expect swift development and implementation of AI.

  1. Brinker TJ, Hekler A, Enk AH, Klode J, Hauschild A, Berking C et al. Deep learning outperformed 136 of 157 dermatologists in a head-to-head dermoscopic melanoma image classification task. Eur J Cancer 2019;113:47-54.

  2. Olsen TG, Jackson BH, Feeser TA, Kent MN, Moad JC, Krishnamurthy S et al. Diagnostic performance of deep learning algorithms applied to three common Diagnoses in dermatopathology. J Pathol Inform 2018;9:32.

  3. Campanella G, Nehal KS, Lee EH, Rossi A, Possum B, Manuel G et al. A deep learning algorithm with high sensitivity for the detection of basal cell carcinoma in mohs surgery frozen sections. J Am Acad Dermatol 2020.

  4. Hekler A, Utikal JS, Enk AH, Berking C, Klode J, Schadendorf D et al. Pathologist-level classification of histopathological melanoma images with deep neural networks. Eur J Cancer 2019;115:79-83.

  5. Hart SN, Flotte W, Norgan AP, Shah KK, Buchan ZR, Mounajjed T et al. Classification of melanocytic lesions in selected and whole-slide images via convolutional neural networks. J Pathol Inform 2019;10:5.

  6. De Logu F, Ugolini F, Maio V, Simi S, Cossu A, Massi D et al. Recognition of cutaneous melanoma on digitized histopathological slides. Front Oncol 2020;10:1559.

  7. Kulkarni PM, Robinson EJ, Sarin Pradhan J, Gartrell-Corrado RD, Rohr BR, Trager MH et al. Deep learning based on standard H&E images of primary melanoma tumors identifies patients at risk for visceral recurrence and death. Clin Cancer Res 2020;26:1126-34.

  8. Ianni JD, Soans RE, Sankarapandian S, Chamarthi RV, Ayyagari D, Olsen TG et al. Tailored for real-World: a whole slide image classification system validated on uncurated multi-site data emulating the prospective pathology workload. Sci Rep 2020;10:3217.

  9. Mattessich S, Tassavor M, Swetter SM , Grant-Kels JM. How I learned to stop worrying and love machine learning. Clin Dermatol 2018;36:777-8.

  10. Jutzi TB, Krieghoff-Henning EI, Holland-Letz T, Utikal JS, Hauschild A, Schadendorf D et al. Artificial intelligence in skin cancer diagnostics: the patients' perspective. Front Med (Lausanne) 2020;7:233.

All content found on Dermatology World Insights and Inquiries, including: text, images, video, audio, or other formats, were created for informational purposes only. The content represents the opinions of the authors and should not be interpreted as the official AAD position on any topic addressed. It is not intended to be a substitute for professional medical advice, diagnosis, or treatment.

DW Insights and Inquiries archive

Explore hundreds of Dermatology World Insights and Inquiries articles by clinical area, specific condition, or medical journal source.

Access archive

All content solely developed by the American Academy of Dermatology

The American Academy of Dermatology gratefully acknowledges the support from Bristol Myers Squibb.

Bristol Myers Squibb Logo