top of page

Artifical Intelligence are dermatologists' days numbered?

 

The level of sophistication shown by artificial intelligence systems may be able to compete with and soon supersede the ability of even the most experienced dermatologists in accurately identifying and diagnosing cancerous lesions. Furthermore, it is envisaged that deep learning algorithms used by such artificial intelligence systems may be able to recognise the most sinister of lesions and hence prioritise treatment in resource limited situations. With the rise of telemedicine and teledermatology, artificial intelligence will also be able to have a greater reach to patients who for various reasons are inaccessible to the surveillance of dermatologists.

 

The diagnostic problem

Despite greater education of the causes of skin cancer, the incidence of malignant melanoma within the United Kingdom has increased by 50% over the last decade and is projected to rise by 7% by the year 2035 (Cancer Research, 2018). Dermoscopy has significantly improved the accuracy of diagnosis by allowing the user to better visualise and thus identify factors in a lesion that are indicative of melanoma such as patterns of colour, irregular vascularity, pseudopods and radial streaming (Walter FM, 2013). Despite this, human-based inclusion criteria, aide memoirs (such as ABCDE and the 7-point checklist) and other methods of pattern recognition still allow atypical types of melanomatous lesions to slip through the net of diagnostic discovery. It can be concluded therefore that the value of dermoscopy depends highly on the presence of features that are illustrative of classical melanomas (Skvara et al, 2005). Computer technology however, is not beholden to the rules that govern human based diagnosis and artificial intelligence software may be able to extrapolate significant information from pixels that were not deemed to be significant and therefore overseen by  the naked eye. This automated analysis of digitalised images shows promise in having an objective approach to examining every fragment of an image for a more in-depth analysis of a lesion.

 

 

The rise of the machines

Machines have had the capacity to learn since the 1950s (Zakhem at al, 2018) but it is only since the turn of the 21st centaury that we have been able to see this technology take off on an exponential basis. Improvements in speech recognition, genomics and drug discovery have all been able to lend themselves to dramatic advancements in artificial intelligence software (Lecun et al, 2015). This technological take off has been due to an unprecedented rise in image availability, improvements in computer technology and an increase in the level of innovation of deep learning algorithms (Lecun et al, 2015). Within the scope of dermatology, machine learning algorithms (also known as convolutional neural networks  -CNN) are able to ‘feed’ computers thousands of images with disease labels, amalgamating a database of dermatological images (Esteva et al, 2017). Using a form of pixilated pattern recognition, machines are then able to amplify certain aspects relevant for lesion identification and disregard other aspects in order to determine what lesion is presented to them and most importantly to decipher whether that lesion is malignant or benign (Esteva et al, 2017).

 

Researchers from the Stanford University artificial intelligence labs were able to demonstrate that a single deep convolutional neural network algorithm was able to differentiate keratinocyte carcinoma from benign seborrheic keratosis and malignant melanoma from benign melanocytic naevi. The CNN was trained using a database of 129,450 clinical images and was able to differentiate the benign lesion from the malignant  as accurately as 21 board certified dermatologists (Esteva et al, 2017). Another study conducted by Han et al, showed how a deep learning algorithm, having analysed 19,398 clinical images of skin lesions was able to accurately diagnose 12 skin diseases including basal cell carcinoma, squamous cell carcinoma and malignant melanoma to a level comparable with 16 dermatologists (Han et al, 2018) and a study performed  by Haenssle et al, showed that the diagnostic ability of a CNN trained using dermascopic images was able to outperform an international group of 58 dermatologists. As well as this, the CNN continued to outperform the diagnostic ability of the dermatologists despite the latter group being given additional clinical information such as patient history (Haenssle et al, 2018).  It can be demonstrated that with a greater supply of images across a wider range of ethnicities and ages CNN will only improve in its diagnostic ability (Han et al, 2018).

 

 

Man against the machine

Much like the machine, experienced dermatologists are able to amalgamate a large quantity of images using their own neural processor and filter out irrelevant material in order to divide skin lesions into two categories: malignant and benign. The current state of affairs in the human diagnosis of skin cancer starts off with the identification of a suspicious lesion, dermascopic review, biopsy and histopathological assessment (Esteva et al, 2017). How can we prove the need for this process and that the dermatologist should maintain their place in the battle of man versus machine?

One cannot emphasise enough the importance of full-body surveillance in the detection of a primary skin cancer as well as in the discovery of possible signs and symptoms of metastasis. Machines, can also fall short due to man-made shortcomings in that they are immediately limited to what the user chooses to display to it and it cannot be taken for granted that the lesion assessed may not be the solitary problem. As well as this, the dermatologist has the advantage of considering clinical history, the evolution of the mole, family history and being able to use tactile perception to assess the texture of the lesion. The gold standard for diagnosing any cancer is biopsy and histopathological assessment (Mogensen and Jemec, 2007). Until the cells are identified, neither man nor machine can be certain that a cancer has not been missed. Man has the ability to access this diagnostic gold standard whereas an algorithm does not.

 

And what of the psychosocial aspect of a cancer diagnosis being delivered by a machine? Would patients fare better having a one-to-one holistic discussion of their diagnosis with their doctor or a machine- automated message of a high probability of cancer. The answer is invariably the former. Patients are also more likely to trust humans as opposed to an algorithm and are also more likely to lose faith in the ability of an algorithm to make an accurate diagnosis if they have seen the algorithm make a mistake (Dietvorst et al, 2014). This could perhaps be due to human nature and wanting to believe that another human being had greater ability to want what’s best for their fellow man rather than a emotionless machine.

 

 

 

Telemedicine and the extension of dermatological care

By mid-2019, it is projected that there will be 2.5 billion smart phone users worldwide (Statista, 2018). Phones now equipped with 12 megapixel cameras, are able to capture clear high-quality images using optical image stabilisation, zoom and wide angled lenses (Ashique et al, 2015). Teledermatology has allowed dermatological services to be delivered to patients living in geographically remote and inaccessible regions. With regards to skin cancer; a lesion is photographed on a smart phone and process by a convolutional neural network where it undergoes multiple layers of mathematical filters that take into consideration factors such as colour, border and irregularity before providing a probability of what the lesion is (Zakhem at al, 2018). This image can then be immediately stored and viewed in real-time or at a later date by dermatologists across the world who are able to deliver diagnostic and treatment recommendations. There is also the added benefit of continued surveillance of skin lesions of patients from the comfort of their own home.

 

Conclusion

Artificial Intelligence, are dermatologists' days numbered?

 

Computational power itself has increased exponentially over the last century and the ability of artificial intelligence systems has grown to what was beyond the expectations of the scientists who coined the technology. As of 2019, deep-learning algorithms are able to function autonomously and self-download both the high-resolution images required for self- teaching as well as the data required to self-update and make improvements in graphic processing. With the ability to continually update itself and download high resolution images, by 2050 we may be at a place completely unforeseen to us now. Mobile CNN and teledermatology continues to make this form of technology ever important in the 21st centaury by their ability to add to a plethora of dermatological images and by expanding the reach of the dermatologist.

 

Artificial intelligence should therefore be used as a valuable clinical tool alongside dermatologists in order to support diagnosis and provide a rapid second opinion, however it must be noted that artificial intelligence software is limited to the information provided to it by the dermatologist. As of 2019, the role of the dermatologist should remain solidified as the primary player in the diagnosis of skin cancer but working in tandem with artificial intelligence (and not in competition with) the field may be elevated to new heights.

 

Word Count: 1491

 

 

References

  • Ashique KT, Kaliyadan F, Aurangabadkar, 2015, Clinical photography in dermatology using smartphones: An overview, Indian Dermatology Online Journal, 6 (3): 158-163

  • Cancer Research UK, Melanoma skin cancer statistics, https://www.cancerresearchuk.org/health-professional/cancer-statistics/statistics-by-cancer-type/melanoma-skin-cancer#heading-Zero, Accessed January 2019

  • Carrera C, Marchetti MA, Dusza SW, Argenziano G, Braun RP, Halpern AC, Jaimes N, Kittler HJ, Malvehy J, Menzies SW, Pellacani G, Puig S, Rabinovitz HS, Scope A. Soyer HP, Stolz W, Hofmann-Wellenhof R, Zalaudek I, Marghoob AA, 2016, Validity and Reliability of Dermoscopic Criteria Used to Differentiate Nevi From Melanoma: A Web-Based International Dermoscopy Society Study,JAMA dermatology, 152(7), 798-806

  • Dietvorst B, Simmons JP, Massey C, 2014, Algorithm Aversion: People Erroneously Avoid Algorithms, After Seeing Them ErrJournal of Experimental Psychology,

  • Esteva A, Kuprel B, Novoa RA, Ko J, Swetter SM, Blau HM, Thrun S, 2017, Dermatologisy- level classification of skin cancer with deep neural networks, Nature 542, 115-118

  • Haenssle HA, Fink C, Schneiderbauer R, Toberer F, Buhl T, Blum A, Kalloo A, Hassen AB, Thomas L , Enk A, Uhlmann L, 2018 Reader study level-I and level-II Groups; Man against machine: diagnostic performance of a deep learning convolutional neural network for dermoscopic melanoma recognition in comparison to 58 dermatologists, Annals of Oncology,29 (1), 1836-1842

  • Han SS, Kim MS, Lim W, Park GH, Park I, Chang SE, 2018, Classification of the Clinical Images for Benign and Malignant Cutaneous Tumors Using a Deep Learning Algorithm, Journal of Investigative Dermatology, 138 (7), 1529-1538

  • LeCun Y, Bengio Y, Hinton G, 2015, Deep Learning, Nature 521, 436-444

  • Mogensen M, Jemec GB, 2007, Diagnosis of Nonmelanoma Skin Cancer/Keratinocyte Carcinoma: A Review of Diagnostic Accuracy of Nonmelanoma Skin Cancer Diagnostic Tests and Technologies, Dermatological Surgery, 33 (10)

  • Skvara H, Teban L, Foebiger M, Binder M, Kittler H, 2005, Limitations of Dermascopy in the Recognition of Melanoma, Archives of Dermatology, 141 (2): 155-160

  • Statista, Number of smartphone users worldwide from 2014 to 2020 (in billions), 2015, https://www.statista.com/statistics/330695/number-of-smartphone-users-worldwide/ Date Accessed: January 2019

  • Walter FM, Prevost T, Hall PN, Vasconcelos J, Burrows NP, Morris H, Kinmoth AL, Emery JD, 2013, Using the 7-point checklist as a diagnostic aid for pigmented skin lesions in general practice: a diagnostic validation study, British Journal of General Practice, 63 (610), 345-353

bottom of page