Background: Cryptococcosis and talaromycosis are known as 'neglected epidemics' due to their high case fatality rates and low concern. Clinically, the skin lesions of the two fungal diseases are similar and easily misdiagnosed. Therefore, this study aims to develop an algorithm to identify cryptococcosis/talaromycosis skin lesions.
Methods: Skin images of tararomiasis and cryptococcosis were collected from published articles and augmented using the Python Imaging Library (PIL). Then, five deep artificial intelligence models, VGG19, MobileNet, InceptionV3, Incept ResNetV2 and DenseNet201, were developed based on the collected datasets using transfer learning technology. Finally, the performance of the models was evaluated using sensitivity, specificity, F1 score, accuracy, AUC and ROC curve.
Results: In total, 159 articles (79 for cryptococcosis and 80 for talaromycosis), including 101 cryptococcosis skin lesion images and 133 talaromycosis skin lesion images, were collected for further mode construction. Five methods showed good performance for prediction but did not yield satisfactory results for all cases. Among them, DenseNet201 performed best in the validation set, followed by InceptionV3. However, InceptionV3 showed the highest sensitivity, accuracy, F1 score and AUC values in the training set, followed by DenseNet201. The specificity of DenseNet201 in the training set is better than that of InceptionV3.
Conclusions: DenseNet201 and InceptionV3 are equivalent to the optimal model in these conditions and can be used in clinical settings as decision support tools for the identification and classification of skin lesions of cryptococcus/talaromycosis.
Keywords: cryptococcosis; deep learning; recognition; talaromycosis.
© 2023 Wiley-VCH GmbH.