Standard of care diagnostic procedure for suspected skin cancer is microscopic examination of hematoxylin & eosin stained tissue by a pathologist. Areas of high inter-pathologist discordance and rising biopsy rates necessitate higher efficiency and diagnostic reproducibility. We present and validate a deep learning system which classifies digitized dermatopathology slides into 4 categories. The system is developed using 5,070 images from a single lab, and tested on an uncurated set of 13,537 images from 3 test labs, using whole slide scanners manufactured by 3 different vendors. The system's use of deep-learning-based confidence scoring as a criterion to consider the result as accurate yields an accuracy of up to 98%, and makes it adoptable in a real-world setting. Without confidence scoring, the system achieved an accuracy of 78%. We anticipate that our deep learning system will serve as a foundation enabling faster diagnosis of skin cancer, identification of cases for specialist review, and targeted diagnostic classifications.Every year in the United States, 12 million skin lesions are biopsied 1 , with over 5 million new skin cancer cases diagnosed 2 . After a skin lesion is biopsied, the tissue is fixed, embedded, sectioned, and stained with hematoxylin and eosin (H&E) on glass slides, ultimately to be examined under microscope by a dermatologist, general pathologist or dermatopathologist who provides a diagnosis for each tissue specimen. Owing to the large variety of over 500 distinct skin pathologies 3 and the severe consequences of a critical misdiagnosis 4 , diagnosis in dermatopathology demands specialized training and education. Although the inter-observer concordance rate in dermatopathology is estimated to be between 90 and 95% 5,6 , there are some distinctions which present frequent disagreement among pathologists, such as in the case of melanoma vs. melanocytic nevi 7-11 . Any system which could improve diagnostic accuracy provides obvious benefits for dermatopathology labs and patients; however, there are substantial benefits also to improving the distribution of pathologists' workloads 12-14 . This can reduce diagnostic turnaround times in several scenarios. For example, when skin biopsies are interpreted initially by a dermatologist or a general pathologist, prior to referral to a dermatopathologist, it can result in a delay of days, sometimes in critical cases. In another common scenario, additional staining is required to identify characteristics of the tissue not captured by standard H&E staining. If those additional stains are not ordered early enough, there can be further delays to diagnosis. An intelligent system to distribute pathology workloads could alleviate some of these bottlenecks in lab workflows. The rise in adoption of digital pathology 1,15 provides an opportunity for the use of deep learning-based methods for closing these gaps in diagnostic reliability and efficiency 16,17 .

Background:Artificial intelligence is advancing at an accelerated pace into clinical applications, providing opportunities for increased efficiency, improved accuracy, and cost savings through computer-aided diagnostics. Dermatopathology, with emphasis on pattern recognition, offers a unique opportunity for testing deep learning algorithms.Aims:This study aims to determine the accuracy of deep learning algorithms to diagnose three common dermatopathology diagnoses.Methods:Whole slide images (WSI) of previously diagnosed nodular basal cell carcinomas (BCCs), dermal nevi, and seborrheic keratoses were annotated for areas of distinct morphology. Unannotated WSIs, consisting of five distractor diagnoses of common neoplastic and inflammatory diagnoses, were included in each training set. A proprietary fully convolutional neural network was developed to train algorithms to classify test images as positive or negative relative to ground truth diagnosis.Results:Artificial intelligence system accurately classified 123/124 (99.45%) BCCs (nodular), 113/114 (99.4%) dermal nevi, and 123/123 (100%) seborrheic keratoses.Conclusions:Artificial intelligence using deep learning algorithms is a potential adjunct to diagnosis and may result in improved workflow efficiencies for dermatopathologists and laboratories.

No abstract

Recent studies have shown an association between obstructive sleep apnea (OSA) and cognitive impairment. This study was done to investigate whether varied levels of cyclical intermittent hypoxia (CIH) differentially affect the microvasculature in the hippocampus, operating as a mechanistic link between OSA and cognitive impairment. We exposed C57BL/6 mice to sham [continuous air, arterial O2 saturation (SaO2 ) 97%], severe CIH to inspired O2 fraction (FiO2 ) = 0.10 (CIH10; SaO2 nadir of 61%), or very severe CIH to FiO2 = 0.05 (CIH5; SaO2 nadir of 37%) for 12 h/day for 2 wk. We quantified capillary length using neurostereology techniques in the dorsal hippocampus and utilized quantitative PCR methods to measure changes in sets of genes related to angiogenesis and to metabolism. Next, we employed immunohistochemistry semiquantification algorithms to quantitate GLUT1 protein on endothelial cells within hippocampal capillaries. Capillary length differed among CIH severity groups (P = 0.013) and demonstrated a linear relationship with CIH severity (P = 0.002). There was a strong association between CIH severity and changes in mRNA for VEGFA (P < 0.0001). Less strong, but nominally significant associations with CIH severity were also observed for ANGPT2 (PANOVA = 0.065, PTREND = 0.040), VEGFR2 (PANOVA = 0.032, PTREND = 0.429), and TIE-2 (PANOVA = 0.006, PTREND = 0.010). We found that the CIH5 group had increased GLUT1 protein relative to sham (P = 0.006) and CIH10 (P = 0.001). There was variation in GLUT1 protein along the microvasculature in different hippocampal subregions. An effect of CIH5 on GLUT1 mRNA was seen (PANOVA = 0.042, PTREND = 0.012). Thus CIH affects the microvasculature in the hippocampus, but consequences depend on CIH severity.

Purpose: To develop a simple and high throughput method for the molecular imaging of proteins at the blood‐brain barrier using immunohistochemistry quantification. Methods: C57B6 male mice, age 4 months, were exposed to one of four different conditions for 12 hours/day × 2 weeks (a) Sham‐‐exposure to constant air‐‐e.g. 21% oxygen (b) cyclical intermittent hypoxia (CIH) from 21% to 12% (c) CIH from 21% to 10% and (d) CIH from 21% to 5%. Immunohistochemistry was then optimized for proteins GLUT1, p‐glycoprotein and CD31 on blood vessels at the bloodbrain barrier. The image data sets of four different conditions were subjected to our image analysis algorithm. Background noise is eliminated using thresholding techniques and blood vessel candidates are delineated using methods based on connected component analysis. Subsequently, relevant features are extracted from the candidate set and are passed through our classification algorithm for capillary identification. We then run a reinforcement algorithm on a subset of images to improve identity classification. This is accomplished by calculating the false positive/negative rate of the subset and feeding these results back to the classification algorithm. We introduce the idea that the gold standard should be computer + human vision. Results: For GLUT1 we analyzed over 5000 images and identified over 400k blood vessels with a false positive rate of 4.1%. For p‐glycoprotein we analyzed over 5000 images and identified over 400k blood vessels with a false positive rate of 5%. Conclusion: Our method provides quantitative measurements of proteins on blood vessels at the blood‐brain barrier in an objective, reproducible and comparable way using standard immunohistochemistry techniques and bright‐field microscopy.