Artificial intelligence should result in better screening, diagnosis and treatment of numerous cancers. Some of the research currently underway is promising, particularly in medical imaging, with already-conclusive results for skin cancer and cervical cancer.
Artificial intelligence, situated somewhere between myth and truth, is providing the collective unconscious much food for thought. If some machines are now clearly outperforming homo sapiens, the “existential risk” as quoted by Stephen Hawking and Elon Musk, is a long way from realisation. There’s a huge gap between self-supervised learning and total machine autonomy.
Recent technological improvements pinpoint, nonetheless, clear progress in healthcare. Artificial intelligence should help usher in a more accurate and, indeed, personalised form of healthcare that will prove invaluable to medical decision-making. Most experts believe, however, that the human factor will predominate, basically on grounds of experience and responsibility. The technology does not yet exist that performs, all by itself, a top-to-toe medical.
Research, prevention, screening, diagnosis, treatment: the scope of IA is almost unlimited. Numerous developments in oncology (a global public health priority) are currently underway. Better still, we are already seeing signs of these advances, especially in medical imaging.
Promising results in skin cancer...
Early diagnosis of cancer is a major challenge impacting the survival of many patients all over the world. The earlier the disease is detected, the more quickly treatment can be commenced and the greater the chances of cure. In this respect, certain research activities are highly promising. One international team, comprising German, American and French scientists, has developed an algorithm 95%-capable of identifying melanoma from simple photographic images*. Of note: the results output by this tool were better than the diagnosis of 58 dermatologists from 17 different countries (87%). The results are even more telling in that they are based on an analysis of 100 rare – as well as complex – cases.
Another baseline study confirms the potential of AI in skin-cancer screening. Dermatologists and engineers from Stanford University in the US have developed a tool capable of distinguishing benign from malignant moles**. Drawn from 130 000 images taken from the Internet, IA managed to identify more than 2 000 skin disorders. The software’s proficiency level is equal to, indeed better than, that of the 21 dermatologists with whose results it was compared.
There is no shortage of public health challenges, particularly in terms of incidence and mortality. According to the International Agency for Research on Cancer (IARC), 232 000 new cases of malignant melanoma are reported every year, resulting in almost 55 000 deaths.
... and in cervical cancer
Artificial Intelligence prowess extends to the early detection of cervical cancer. American researchers have developed an algorithm capable of identifying precancerous lesions on photographs, with a success rate of 91%***. For information: the tool was applied to a bank of 60 000 images showing a mix of healthy and pathological cervixes. The results are spectacular, since the machine’s level of accuracy was substantially higher than that obtained by traditional methods, namely, readings of colposcopy images taken by experts (69%) or the cytology pap smear (71%).
Here again, the challenge will be to facilitate the early detection of this cancer, in particular in developing countries, which report the highest mortality rate. The WHO reports cervical cancer as the fourth most commonly diagnosed cancer in the world, with 570 000 cases reported in 2018. In the developed countries, the routine screening of populations at risk and the multiple vaccination campaigns introduced to combat the different forms of HPV have gone a long way to reducing cervical cancer mortality rates. Artificial intelligence could yet increase this welcome trend.
Multiple challenges
Essentially multi-faceted, artificial intelligence is a source of opportunities galore. Predictive analysis could, for example, prevent even more cancers, by better anticipating behavioural risks. It could also help optimise clinical research protocols. Experts see AI as a means of consolidating the feasibility of studies, of improving the predictability of the therapeutic effects and also of better measuring the usefulness of the solutions implemented. In reality, the future of AI will depend greatly on the eventual use made of health data, the drivers, in this case, of the AI. As is widely acknowledged, advances in machine learning will depend mainly on the quality of the information input manually.
The recourse to artificial intelligence – which is cheaper, faster and often more effective – should help reduce the prevalence and incidence of many cancers. Although most progress has been made in imaging, other research projects are underway, such as a more precise assessment of the aggressivity of the tumour or even a more accurate targeting of patients eligible for certain therapies (see diagram). The biggest challenge will be to support the most promising developments, in the general interest. A life-saving approach in which Candriam has every intention of participating, by identifying, selecting and helping the most dynamic companies in this now-strategic sector.
Immunotherapy: better profiling of eligible patients in the pipeline?
The news did not leave the international scientific community unmoved. French researchers have developed and trialled an AI algorithm capable of predicting a patient’s positive reaction to an immunotherapy from scanned images****, a technique considered a viable alternative to a biopsy, being less invasive and, above all, less risky (depending on the location of the tumour). It could also lessen the cost of expensive treatment. More eligible patients can also be treated, with significant savings for healthcare systems. As it stands, the technology is capable of correctly determining the immune profile of a tumour in almost 60% of cases. Research studies continue to improve already highly promising results.
(*) "Man against machine: diagnostic performance of a deep learning convolutional neural network for dermoscopic melanoma recognition in comparison to
(**) "Dermatologist-level classification of skin cancer with deep neural networks": Nature (February 2017).
(***) "An observational study of deep learning and automated evaluation of cervical images for cancer screening": Journal of the National Cancer Institute (January 2019).
(****) "A radiomics approach to assess tumour-infiltrating CD8 cells and response to anti-PD-1 or anti-PD-L1 immunotherapy: an imaging biomarker, retrospective multicohort study": The Lancet (August 2018).