|Description||This page gives a brief introduction and overview on neural networkThis glossary term has not yet been described. based automated diagnosisis the identification of the nature and cause of a certain phenomenon. Diagnosis is used in many different disciplines with variations in the use of logic, analytics, and experience to determine "cause and effect". In systems engineering and computer science, it is typically used to determine the causes of symptoms, mitigations, and solutions of skin lesions with dermatoscopyThe examination of [skin lesions] with a 'dermatoscope'. This traditionally consists of a magnifier (typically x10), a non-polarised light source, a transparent plate and a liquid medium between the instrument and the skin, and allows inspection of skin lesions unobstructed by skin surface reflections. Modern dermatoscopes dispense with the use of liquid medium and instead use polarised light to cancel out skin surface reflections..|
|Responsible author||Tschandl → send e-mail|
|Status update||June 24, 2018|
|Status by||Ralph P. Braun|
Neural networks, Pytorch, Tensorflow, Keras, Mxnet, Chainer, Caffe, Deep learning, Dermoscopy, Dermatoscopy, Diagnosis, Skin cancer, Melanoma Automated diagnosis – cite! Automated diagnosis (message) Automated diagnosis – participate!
Machine learning techniques, and neural networksThis glossary term has not yet been described. in particular, for diagnosing melanomaThis glossary term has not yet been described. have been in focus of research for more than two decades in dermatology. A systematic review as early as 2009 has shown that automated classifiers can have a comparable diagnostic odds ratio to physicians. Systems reviewed in this case commonly relied on manual or semi-automatic feature extraction of digital imagesA representation of a person, animal or thing, photographed, painted or otherwise made visible.. Unprecedented image classificationis a general process related to categorization, the process in which ideas and objects are recognized, differentiated, and understood. accuracy through so-called deep convolutional neural networks (CNN), which don't need a feature extraction step, has regained interest in automated analyses in the medical domain.
Comparison to human accuracy
Several studies found human level performance of such state-of-the-art CNNs to dermatologists for melanoma diagnosis or multi-class accuracy. Similar rates were found for classificatin of close-up images.
CNNs need a large number of images during training, are thus highly dependant on datasets providing those - and are currently limited to detecting diagnoses covered in them. At the moment the largest publicly available datasets are:
- HAM10000, available for downloadThis glossary term has not yet been described. at the Harvard Dataverse and ISIC-Archive
- EDRA Interactive Atlas of Dermoscopy
- PH2 Database
The future central hub for any public dataset is probably the ISIC-Archive, where already datasets from Memorial Sloan Kettering Cancer Center, the SONIC study and othersThis glossary term has not yet been described.This glossary term has not yet been described. are available in a structured way.
Although published accuracy rates for neural network classifiers - similar to studies >10 years ago - seem promising, researchers are raising concerns whether neural network classification can live up to its current promise to "solve" automated melanoma recognition. This is due to many reasons, some exemplarily shown below:
- Instability: Current neural networks are highly sensitive to small image perturbations invisible to the human eye, shown by adversarial attacks against medical images. But also more evident differences in images can have an impact: E.g. if only melanoma images in the training dataset are photographed with skin markings or rulers, a network may also "diagnose" images of angiomas as melanomas if depicted alongside rulers and skin markings.
- Explainability: Although every parameter of a neural network is known, explainability of those systems is impeded by the vast number of parameters. One approach for intuitively explaining neural network decisions is by simply providing similar images with "content based image retrieval" (CBIR). A free example application provided by the ViDIR Group (Medical University of Vienna) can be used at ypsono.com.
- Evaluation: Classic evaluation metrics of studies and machine learning challenges show comparability to human accuracy, but it is debateable if they provide valuable insight for performance and clinical applicability.
- Real world application: While recent studies show users accept automated classifiers in principleThis glossary term has not yet been described., more in depth early studies showed that only non-experts, and only in ambiguous cases, would change their decision based on an automated system. It is completely unclear whether published metrics accurately reflect this population and cases, and whether automated decisions in those cases can provide beneficial and safe outcomes. An underestimated danger is also that an automated system can of course not diagnose a melanoma that is not being photographed at all, which was in fact the case in a prospective clinica trial in Vienna.
Apart from image analysis and classification of dermatoscopic images, several other techniques try to solve melanoma diagnosis (list non-exhaustive):
- Multispectral image analysis
- Raman spectroscopy 
- Electrical Impedance Spectroscopy
- Genomic analysis through adhesive tape
They all try to provide a binary output with little to no user interaction. Albeit accuracy rates have been reported high for some of those, it is still unclear if and how they will provide meaningful input in clinical practice.