Biased Algorithms in Melanoma Diagnoses

Infiltrating the age of digital expansion, but also the peak of medical development, the age-old persistence of cancers is still set at a record high – affecting 1 in 2 people at some stage in their lives - melanoma makes up 5% of all cancer cases in the UK. People of colour are up to 5 times more likely to die from skin cancer than Caucasians [Hugh M. Gloster Jr. MD, 2006], despite lower rates of occurrence. With rapid recent developments in the machine learning medical sector, how and why do computer algorithms possess racial biases? There were 331,647 newly diagnosed cases of melanoma in 2022, and 58,645 deaths [Freddie Bray, 2024].  In a 2006 Study of ‘Comparison of Stage at Diagnosis of Melanoma Among Hispanic, Black and White Patients in Miami-Dade County, Florida’, [Shasa Hu et al] found that out of 1690 melanoma cases, only 2% were among black patients, of which 52% had a late-stage diagnosis compared to the 70% being white patients, and just 16% of them receiving a late-stage diagnosis. Black people, although less likely to contract melanoma, are around 35% more likely to receive a late-stage diagnosis, leading to higher rates of mortality. In addition, a data shortage in photographs and cell visualisations, has led to fewer and less accurate machine learning models. It is vital to improve research and awareness for diagnosing black patients at an earlier stage, to reduce the mortality of melanoma diagnoses. Artificial intelligence is a tool that many sectors are quickly becoming reliant upon. In medicine, [Seung Seog Han et al, 2018] AI was taught to classify benign and malignant tumours. There was an 81% accuracy rate on the Asan training set, equivalent to the accuracy of 16 dermatologists. The comparability of deep learning to specialists in their field shows that machines in this digital age are extremely intelligent - these technologies can be crucially used to diagnose those of underrepresented minorities. In addition, this study also found that the accuracy yield was higher when data sets focused on specific skin tones. When using data gathered from Caucasians to diagnose other races, there was an accuracy of only 56% - further emphasising the importance of including variety in research. With deep learning, the accuracy of algorithms has reduced the need for painful and invasive biopsies which is the common method for skin cancer detection. In a paper published by Mehwish Dildar on convolutional neural network (CNN)-based skin cancer techniques, they were able to use deep learning to teach an algorithm how to recognise and classify skin legions. By exposing the program to past skin cancer cases, it was able to effectively use computer vision to classify images – yielding an 85.5% accuracy in melanoma classification. This is a huge advancement for the future of deep learning in medicine, as it shows that machine learning can help to improve future healthcare provisions and efficiency. By using a CNN architecture for melanoma detection, algorithms must be able to process the image into a form of digital data and code. To do this, images are sized to a standardised resolution, but they retain extra detail in the dimensions of the legion, stored in a 3D array. Deep learning algorithms can process which legions are possibly cancerous by finding common threads in places where syntactic code is similar. These algorithmic advancements would cyclically help to diversify the information available to machine learning algorithms and decrease cases of late-stage melanoma diagnoses. Conversely, detailed research lacks minority representation, with Caucasians usually being at the center of funded research. Furthermore, an algorithm that can only access certain data types can only learn to the extent of its information access. A dermatological questionnaire conducted by M. Lyman in 2016 proved that GPs in England misdiagnosed melanoma up to 5 times more often in black patients. When exposed to such cases, machine algorithms struggle equally to diagnose black patients. In the digital age, advancements in machine learning present life-saving opportunities. We must ensure that they save lives regardless of skin colour. Opportunities to do this are everywhere – from funding research to increased documentation of individual melanoma cases. At the cusp of machine learning domination, this is more crucial than ever.