User login
a proof-of-concept study reported.
Early diagnosis of infantile hemangiomas “is essential, as there is a narrow window of opportunity to treat high-risk lesions,” April J. Zhang, MD, and coauthors noted in the study. “AI algorithms optimized for image classification through use of convolutional neural networks have been widely utilized to classify lesions in which images are readily standardized, such as skin cancers and onychomycosis.”
The results were published in Pediatric Dermatology.
Dr. Zhang, of the department of dermatology at the Medical College of Wisconsin, Milwaukee, and colleagues trained a convoluted neural network to diagnose infantile hemangiomas based on clinical images from pediatric dermatology patients treated at Children’s Wisconsin between 2002 and 2019.
They used Microsoft’s ResNet-50, a publicly available network architecture, to train a binary infantile hemangioma classifier to group images as infantile hemangiomas or non–infantile hemangiomas. The team randomly split data from the model into training, validation, and test groups.
The preliminary data set contained 14,811 images, about half of which were facial lesions. The training group of images achieved an accuracy of 61.5%. Next, Dr. Zhang and colleagues limited the data set to facial-only lesions and removed poor-quality images, which left 5,834 images in the final data set: 4,110 infantile hemangiomas and 1,724 non–infantile hemangiomas. This model achieved an overall accuracy of 91.7%, with a sensitivity of 93% and a specificity of 90.5%.
“Our study is the first to demonstrate the applicability of AI in the pediatric dermatology population,” the authors wrote. “With current nationwide shortages in pediatric dermatologists, AI has the potential to improve patient access and outcomes through enhanced rapid diagnostic capabilities.”
They acknowledged certain limitations of the study, including a data set with greater numbers of infantile hemangiomas, compared with non–infantile hemangiomas.
“Random oversampling of the non–infantile hemangioma data set was used to combat this but may lead to model overfitting, where a model performs well on its training data but is unable to generalize to new data,” they wrote. “As infantile hemangiomas are rarely biopsied, expert clinical diagnoses were used as the gold standard without pathologic confirmation.”
The authors reported having no financial disclosures.
a proof-of-concept study reported.
Early diagnosis of infantile hemangiomas “is essential, as there is a narrow window of opportunity to treat high-risk lesions,” April J. Zhang, MD, and coauthors noted in the study. “AI algorithms optimized for image classification through use of convolutional neural networks have been widely utilized to classify lesions in which images are readily standardized, such as skin cancers and onychomycosis.”
The results were published in Pediatric Dermatology.
Dr. Zhang, of the department of dermatology at the Medical College of Wisconsin, Milwaukee, and colleagues trained a convoluted neural network to diagnose infantile hemangiomas based on clinical images from pediatric dermatology patients treated at Children’s Wisconsin between 2002 and 2019.
They used Microsoft’s ResNet-50, a publicly available network architecture, to train a binary infantile hemangioma classifier to group images as infantile hemangiomas or non–infantile hemangiomas. The team randomly split data from the model into training, validation, and test groups.
The preliminary data set contained 14,811 images, about half of which were facial lesions. The training group of images achieved an accuracy of 61.5%. Next, Dr. Zhang and colleagues limited the data set to facial-only lesions and removed poor-quality images, which left 5,834 images in the final data set: 4,110 infantile hemangiomas and 1,724 non–infantile hemangiomas. This model achieved an overall accuracy of 91.7%, with a sensitivity of 93% and a specificity of 90.5%.
“Our study is the first to demonstrate the applicability of AI in the pediatric dermatology population,” the authors wrote. “With current nationwide shortages in pediatric dermatologists, AI has the potential to improve patient access and outcomes through enhanced rapid diagnostic capabilities.”
They acknowledged certain limitations of the study, including a data set with greater numbers of infantile hemangiomas, compared with non–infantile hemangiomas.
“Random oversampling of the non–infantile hemangioma data set was used to combat this but may lead to model overfitting, where a model performs well on its training data but is unable to generalize to new data,” they wrote. “As infantile hemangiomas are rarely biopsied, expert clinical diagnoses were used as the gold standard without pathologic confirmation.”
The authors reported having no financial disclosures.
a proof-of-concept study reported.
Early diagnosis of infantile hemangiomas “is essential, as there is a narrow window of opportunity to treat high-risk lesions,” April J. Zhang, MD, and coauthors noted in the study. “AI algorithms optimized for image classification through use of convolutional neural networks have been widely utilized to classify lesions in which images are readily standardized, such as skin cancers and onychomycosis.”
The results were published in Pediatric Dermatology.
Dr. Zhang, of the department of dermatology at the Medical College of Wisconsin, Milwaukee, and colleagues trained a convoluted neural network to diagnose infantile hemangiomas based on clinical images from pediatric dermatology patients treated at Children’s Wisconsin between 2002 and 2019.
They used Microsoft’s ResNet-50, a publicly available network architecture, to train a binary infantile hemangioma classifier to group images as infantile hemangiomas or non–infantile hemangiomas. The team randomly split data from the model into training, validation, and test groups.
The preliminary data set contained 14,811 images, about half of which were facial lesions. The training group of images achieved an accuracy of 61.5%. Next, Dr. Zhang and colleagues limited the data set to facial-only lesions and removed poor-quality images, which left 5,834 images in the final data set: 4,110 infantile hemangiomas and 1,724 non–infantile hemangiomas. This model achieved an overall accuracy of 91.7%, with a sensitivity of 93% and a specificity of 90.5%.
“Our study is the first to demonstrate the applicability of AI in the pediatric dermatology population,” the authors wrote. “With current nationwide shortages in pediatric dermatologists, AI has the potential to improve patient access and outcomes through enhanced rapid diagnostic capabilities.”
They acknowledged certain limitations of the study, including a data set with greater numbers of infantile hemangiomas, compared with non–infantile hemangiomas.
“Random oversampling of the non–infantile hemangioma data set was used to combat this but may lead to model overfitting, where a model performs well on its training data but is unable to generalize to new data,” they wrote. “As infantile hemangiomas are rarely biopsied, expert clinical diagnoses were used as the gold standard without pathologic confirmation.”
The authors reported having no financial disclosures.
FROM PEDIATRIC DERMATOLOGY