In a recent study published in the Medicine in Novel Technology and Devices Journal, researchers used a large dataset comprising images of skin lesions from monkeypox (mpox) patients to develop a machine-learning-based detection tool to detect mpox.
Study: Deep learning based detection of monkeypox virus using skin lesion images. Image Credit: sulit.photos/Shutterstock.com
Mpox is a zoonotic, systemic disease caused by the monkeypox virus (MPV), which belongs to the Orthopoxvirus genus of the Poxviridae family.
Until early 2022, the disease had been endemic to west and central Africa, but as of late 2022, monkeypox cases have been reported from more than a hundred countries outside of the endemic region, and this recent spread of MPV to North America and Europe is being considered a global outbreak.
Along with fever, headaches, muscle pain, and swelling of lymph nodes, the disease also causes rashes and lesions on the palms, soles, and face and in the mucous membranes of the mouth and genital regions.
The rashes begin on the soles and palms, spread to the eyes, genitals, and mouth, and usually progress from the flat or macule form to firm, raised lesions called papules which finally get pus filled to form pustules.
The current standard method for detecting monkeypox is using polymerase chain reaction (PCR) tests, which can often be inconclusive due to the short duration the virus remains in the body and or inaccessible in rural and remote areas.
However, artificial intelligence and machine learning methods provide faster and more accessible disease diagnosis methods.
About the study
The present study developed a model based on deep learning methods to detect mpox using skin lesion images taken on regular smartphone cameras. The study aimed to use various deep learning methods, including AlexNet and GoogLeNet, to detect mpox accurately.
They also compared the performance metrics of other machine learning models used to diagnose mpox in terms of accuracy, recall, precision, and f1-score.
The training dataset comprised 228 images, of which 102 were of mpox, and the remaining 126 were of measles and chickenpox lesions. Various augmentation methods such as translation, rotation, shear, reflection, hue, contrast, brightness, saturation, and scaling were used to increase the dataset, which consisted of 1,428 images of mpox lesions and 1,764 photos of other lesions.
The deep neural networks were trained using the training image dataset in Deep Network Designer run on MATLAB 2022. Pilot runs were conducted for several neural networks, including Places365-GoogleNet, GoogLeNet, AlexNet, SqueezeNet, Vision Transformer, and ResNet-18.
The results reported that of all the tested neural networks, the results from ResNet-18 had the highest accuracy (99.49%).
The researchers believe that ResNet-18 performed with better accuracy than Places365-GoogleNet, Squeezenet, and GoogLeNet due to its effective and straightforward architecture, which allowed it to learn the complex features of the detection method without numerous inputs. ResNet-18 also has fewer convolutional layers than the other neural networks and makes lower demands on computer memory.
The Vision Transformer model was used as an alternative to the conventional neural network models, and it was found to perform poorly in comparison to the neural network models when using similar training and validation hyperparameters.
This difference in performance could be due to vision-transforming models requiring a large training dataset courtesy of their numerous parameters.
Deep learning methods in medicine provide faster and more accurate testing options. They can efficiently filter large amounts of patient data without compromising accuracy and time.
Furthermore, the resource efficiency and the lack of heavy or expensive equipment make it an ideal mpox detection method in various healthcare settings and clinics in various regions.
To summarize, the researchers used a large dataset of mpox lesions and lesions from measles and chickenpox to train various neural networks to detect mpox cases from images taken on easily accessible smartphone cameras.
Overall, the findings indicated that the neural network model ResNet-18 performed the best, with an accuracy of 99.49%.
Furthermore, with other techniques, such as Locally Interpretable Model-agnostic Explanations (LIME), healthcare professionals can potentially use this method to detect mpox and visually interpret the predictions based on the neural network model results.