RESUMO
Background@#Ultrasonography is an effective noninvasive imaging modality for the diagnosis of subcutaneous masses. To date, few studies have reported skin ultrasonography using deep convolutional neural networks (DCNNs).We investigated the accuracy of DCNNs for the diagnosis of epidermal cysts, lipomas, and other subcutaneous masses. @*Objective@#The purpose of this study was to evaluate whether DCNNs could diagnose subcutaneous masses with ultrasonographic images at level of competence comparable to dermatologists. @*Methods@#We created a dataset of 1,361 skin ultrasonography images obtained from 202 patients diagnosed with epidermal cysts, lipomas, and other subcutaneous masses, to train the DCNNs using ResNet18. Performance was compared with another set of 93 ultrasonographic images (24 epidermal cysts, 25 lipomas, and 44 other subcutaneous masses) from open-access articles. @*Results@#The DCNNs yielded 87.10% classification accuracy and 86.10% F1-scores. The area under the curve, sensitivity, and specificity were 0.92 (95% confidence interval [CI] 0.86∼0.98), 75.00%, and 98.55% for epidermal cysts; 0.93 (95% CI 0.88∼0.98), 80.00%, and 94.12% for lipomas; and 0.97 (95% CI 0.93∼1.00), 97.73%, and 85.71% for other subcutaneous masses, respectively. Analysis using gradient-weighted class activation mapping revealed that the DCNNs could detect specific ultrasonographic findings of epidermal cysts and lipomas. @*Conclusion@#We propose that DCNNs combined with ultrasonography may aid in the diagnosis of subcutaneous masses in outpatient settings.
RESUMO
Background@#Ultrasonography is an effective noninvasive imaging modality for the diagnosis of subcutaneous masses. To date, few studies have reported skin ultrasonography using deep convolutional neural networks (DCNNs).We investigated the accuracy of DCNNs for the diagnosis of epidermal cysts, lipomas, and other subcutaneous masses. @*Objective@#The purpose of this study was to evaluate whether DCNNs could diagnose subcutaneous masses with ultrasonographic images at level of competence comparable to dermatologists. @*Methods@#We created a dataset of 1,361 skin ultrasonography images obtained from 202 patients diagnosed with epidermal cysts, lipomas, and other subcutaneous masses, to train the DCNNs using ResNet18. Performance was compared with another set of 93 ultrasonographic images (24 epidermal cysts, 25 lipomas, and 44 other subcutaneous masses) from open-access articles. @*Results@#The DCNNs yielded 87.10% classification accuracy and 86.10% F1-scores. The area under the curve, sensitivity, and specificity were 0.92 (95% confidence interval [CI] 0.86∼0.98), 75.00%, and 98.55% for epidermal cysts; 0.93 (95% CI 0.88∼0.98), 80.00%, and 94.12% for lipomas; and 0.97 (95% CI 0.93∼1.00), 97.73%, and 85.71% for other subcutaneous masses, respectively. Analysis using gradient-weighted class activation mapping revealed that the DCNNs could detect specific ultrasonographic findings of epidermal cysts and lipomas. @*Conclusion@#We propose that DCNNs combined with ultrasonography may aid in the diagnosis of subcutaneous masses in outpatient settings.