Use of artificial intelligence in skin cancer diagnosis and management

Artificial intelligence is a branch of computer science that, in broad terms, deals with either decision making or classification. The aim of artificial intelligence is to surpass human cognitive functioning such that automated decisions can be made. Machine learning — an application of artificial intelligence — is commonly used in image recognition. In general, the machine, or algorithm, learns from exposure to a large dataset. Once learning has taken place, the algorithm can be applied to unseen data. The potential advantages of this approach in health care are clear: machines can learn from very large datasets in relatively short time frames and can apply themselves to new data without fatigue or intra‐observer replication error.

Machine learning has recently demonstrated remarkable performance in image‐based diagnosis across various medical fields, including ophthalmology, radiology, pathology and dermatology. In dermatology, the primary focus has been on developing machine learning systems that facilitate classification and decision support for skin cancer management. Skin cancer (including melanocytic and keratinocytic malignancy) is the most common cancer in Australia and among Caucasian populations worldwide. Melanoma is responsible for the majority of skin cancer deaths in Australia and has various presentations.1,2 While dermoscopy has improved the accuracy of melanoma diagnosis, significant variability occurs and is largely a function of clinical expertise. Recent studies show that machine learning algorithms have the potential to surpass the diagnostic performance of experts, and the challenge now is how to implement this new technology safely into clinical practice.

Although there are a number of machine learning algorithms that could be used in the dermatology setting, convolutional neural networks (CNNs) are the most promising. This is largely because they learn from data without any feature specification, and they are known to exhibit superior performance for image recognition in comparison with other machine learning algorithms.3 The aim of the CNN is to generalise its previously learned knowledge on unseen images beyond the training dataset. There are numerous parameters within a CNN that can be tweaked to maximise algorithm performance. Most of these parameters are adjusted automatically by the algorithm, without user input. Therefore, very little can be known, in principle, about why and how the algorithm reaches any particular decision. Currently, there are efforts underway to reduce the “black box” effect of CNNs. Some commercial software programs coupled to imaging devices will provide the user with comparable lesions to justify the algorithm's output and improve transparency. However, this retrieval system may fail for rare or unseen cases and does not provide a decision‐making process. While the black box phenomenon remains, there are two potentially negative implications for clinical practice: first, clinicians may have difficulty upskilling by following the algorithms’ outputs; and second, there exists the potential for deskilling and underperforming due to an over‐reliance on technology.4,5 The effect of a faulty system has been explored by manipulating a previously trusted algorithm to generate incorrect classifications and found that doctors of all experience levels were susceptible to being misled by the recommendation.5

Algorithm performance is dependent on both the size and quality of the training image dataset and on whether the algorithm is used in situations for which it was intended. Depending on the training set, the device may be limited in its ability to diagnose specific lesions (eg, non‐pigmented), or lesions in certain skin types (eg, darker skin) or sites (eg, scalp or acral). Retrospective image databases used to train algorithms may be associated with bias. In addition, artefacts (eg, hair, dermoscopic gel, air bubbles, rulers, pen markings, reflections) can distract from key features. However, if a CNN is trained on a large enough cohort, it can learn to deal with potential artefacts. Nonetheless, unbiased lesion selection and standardised image capture would invariably improve algorithm performance, and recent advances in three‐dimensional (3D) imaging modalities will enable this.6

Several studies have now shown that CNNs trained on retrospective image data collected at a single time point are capable of classifying skin cancer with sensitivities and specificities equal or superior to that of dermatologists (Box 1),5,7,8,9,11 and clinicians with less experience gain most from AI support under experimental conditions.5 Hypomelanotic and acral melanoma can be more challenging to diagnose clinically,1 and this could potentially present a challenge for automated classification. However, CNNs have achieved greater accuracy for hypopigmented and acral lesions in comparison with human experts, at least in silica.9,11 In addition to clinical images, CNNs have been applied to histopathological images of melanoma and benign naevi with promising results.10

The ground truth for lesion diagnosis

The gold standard for melanoma diagnosis is histopathological assessment. However, there exists significant inter‐ and intra‐observer variability in histological diagnostic labels attributed to atypical melanocytic lesions.12 The existence of such variability in diagnoses poses the dilemma of whether the CNN has learnt from the correct set of diagnoses. Consensus diagnoses, if practical, may help overcome this problem. Molecular biomarkers may assist in establishing a diagnosis13 and identifying high risk biology,14 but they require extensive validation before clinical use. Pathologists and clinicians also rely on metadata (age, personal and family history, lesion symptoms, recent change), which may influence diagnostic likelihoods. Importantly, it is possible to incorporate different data types, including metadata, sequential image data coupled with histopathology, to train future CNN algorithms and improve diagnostic discrimination of borderline lesions (Box 2).

Use of artificial intelligence for melanoma screening

It is well known that the incidence of invasive melanoma in Australia has increased over the past 40 years. In addition, there has been a striking increase in incidence of in situ melanoma over the past decade, from 32 cases per 100 000 population in 2004 to 80 per 100 000 population in 2019, with age‐standardised mortality remaining fairly stable.2 The potential causes for the increase in incidence are complex, and involve a true increase, driven by poor sun exposure practices of individuals born before the SunSmart era, combined with increased awareness, excessive screening, and overdiagnosis. It has recently been estimated that 54% of melanomas (15% of invasive melanomas) are overdiagnosed.15

Artificial intelligence‐assisted targeted screening of high risk individuals is likely to be a more effective strategy to save lives than the current opportunistic approach. With sequential whole‐body image datasets linked to metadata, molecular biomarkers and clinical outcomes, our ability to identify lesions associated with sinister biological potential will improve (Box 2), thereby reducing unnecessary biopsies, minimising overdiagnosis and other potential harms associated with screening.

Use of artificial intelligence in clinical practice

There are advantages and disadvantages of introducing artificial intelligence at different points in the patient care pathway.16 An artificial intelligence system used as a triaging tool before clinician assessment would enable automated risk stratification of individuals and/or lesions (Box 2). This approach could dramatically improve clinician workload and timely access to specialist care for people requiring urgent attention. Alternatively, artificial intelligence consulted following an examination by the clinician may act as a second opinion to improve diagnostic sensitivity and reduce unnecessary biopsies.5 The latter is more closely aligned with current clinical workflows and therefore likely to be preferred while the field matures. There is potential for over‐reliance on artificial intelligence systems in both scenarios.

A secondary support system may provide the clinician with a diagnosis or a management decision. Doctors are more likely to change their minds if they are uncertain of a diagnosis and an algorithm provides a conflicting result.5 It is thus important to consider how an algorithm might convey uncertainty to avoid false guidance. For example, a decision‐support output (eg, excise, monitor or reassure) avoids the diagnostic dilemma of differentiating between melanoma and dysplastic naevi. However, the problem is complex and arguments exist as to why, in many situations, a diagnostic probability output might be more desirable.

Safe implementation of new technologies

The Therapeutic Goods Administration (TGA) has developed an action plan to improve the processes by which new devices are approved for use in Australia, strengthen monitoring and follow‐up, and provide more information to consumers about the devices they use.17 International collaborations also exist with groups, such as the International Medical Device Regulators Forum, to establish better processes for medical device regulation globally. If software is classified as a medical device (ie, it is intended for diagnosis, prevention, monitoring, treatment or alleviation of disease), it must be registered on the Australian Register of Therapeutic Goods following TGA approval and before distribution within Australia.

Consumers and clinicians need to be aware of the intended use of an application or device. There are several smartphone applications available to the general public, with functionality ranging from education to monitoring and tracking to skin lesion classification. Some of these provide skin lesion risk assessment, although they may state that they are not intended to be used as a diagnostic device. There is concern that, if this is not immediately obvious to the consumer, unregistered applications may be used in lieu of seeking medical advice. Unsupervised consumer‐operated diagnostic devices would require careful testing before they can be recommended.

Conclusion

As clinicians, we need to be aware of the limitations of any diagnostic tool and interpret outputs accordingly. Although the performance of artificial intelligence to date is promising, it remains to be seen how diagnostic devices in dermatology will influence decision making in the clinic and affect patient outcomes. Regardless of the specialty, any new technologies need to be rigorously tested before implementation and monitored after implementation. Ultimately, responsibility for patient care remains with the clinician and, as such, a high level of clinical acumen must be maintained. Nonetheless, artificial intelligence in dermatology is primed to become a powerful tool in skin cancer assessment.

Box 2 – Incorporation of different data types to train future convolutional neural network (CNN) algorithms and improve diagnostic discrimination of borderline lesions

---

AI = artificial intelligence.