Can we already trust AI to diagnose cancer?

Published January 12 2023

What will tomorrow's place be for artificial intelligence (AI) in medicine? Could the diagnostic problems no longer be solved by a Dr. House but by his digital alter ego? In fact, AI is already surpassing the eye in detecting certain tumors from radiological images (mammograms, MRIs)… Which has led some to say that computers will soon replace human experts.

But, contrary to these predictions, the radiologist did not disappear : on the contrary, an unexpected “collaboration” took place between him and the machine which was to take his place. The first working to channel the capacities and strengths of the second in order to improve interpretation and diagnosis for the benefit of patients.

This question of aid to the correct diagnosis is central, and is worth both in psychiatry where AI is also taking its first steps than in oncology… In pathological anatomy, either "the examination of organs, tissues or cells to identify and analyze abnormalities related to a disease (cancer, etc.)", the prospects and promises are enormous.

Is AI already capable of such analyses? Could it turn out to be more efficient than the human expert?

Misunderstandings and confusion abound, and it is important to understand why. It is this point that we propose to you here.

What allowed the first steps of “digital pathology”

For the AI, as for any human specialist, the diagnosis is based, among other things, on an object as simple as it is essential: the glass slides on which the pathologist places a very thin "slice" of the tissue to be analyzed (lung, liver , etc.), in order to observe it under the microscope.

Through this microscopic analysis, the pathologist can identify different types of cells, compare their shapes or even their spatial organization (architecture) to identify abnormal clusters – tumors for example.

The mass digitization of these slides paved the way for the use of AI in pathological anatomy. The advent of adapted scanners allows, in a growing number of hospitals, the acquisition and storage of microscopy slides in digital form. The original slides are however kept… which will not necessarily be possible for all of their digitized versions, due to the cost of storage.

This procedure, which paves the way for “digital pathology”, has made it possible to work on algorithms intended to carry out their analysis in an automated way. With the objective that AI can assist the pathologist in his diagnosis. It is also useful for ergonomic reasons and to save time.

Glass slides are traditionally observed under a microscope. They can now be digitized for study on a computer screen. This also allows them to be transmitted to artificial neural networks.
DR, Supplied by the author

But like the human, the machine (most often artificial neural networks) must be trained. First, she must be able to “look” at the blades and understand what it is about. This analysis uses pattern recognition technology as the basic technique.

Second, it must be able to interpret what it “sees”. AI is based on the notion of learning and the ability to infer, i.e. to transfer the knowledge acquired during its formation and training to other situations, comparable but not similar: for example , recognize a breast cancer lymph node micrometastasis (cluster of a few tumor cells that may go unnoticed) by having previously seen other images of metastases.

It should be noted that digitized slides contain many more pixels than radiological images and contain thousands of cells – they are therefore particularly rich in information that algorithms could exploit.

[Nearly 80 readers trust The Conversation newsletter to better understand the world's major issues. Subscribe today]

A fast and reliable digital assistant…

Current research and trials show that AI could ultimately be relevant in several areas:

automation of the most repetitive and subjective activities,
aids in tumor detection, aggressiveness assessment and subtyping,
counting of tumor cells, especially those in division (mitoses),
evaluation of the intensity of the immune response (number of lymphocytes attacking the tumour).

The interests are multiple: to give time back to the human pathologist so that he can devote himself to the most complex tasks where the human added value is real, to make the final diagnosis faster and more reliable. And, importantly in science, the results of AI analyzes are generally reproducible.

We can already identify concrete cases where the contribution of AI is relevant:

Breast cancer detection: algorithms are more efficient than the pathologist in detecting detection of micrometastases in the lymph nodes of the axillary hollow.
Breast cancer prognosis assessment: artificial neural networks effectively identify cell markings made using specific antibodies (immunohistochemistry technique). In breast cancer, quantifying the expression of the HER2 protein in tumor cells makes it possible to assess the prognosis of the disease and the response to certain drugs – this protein stimulates cancerous progression. Computer-aided diagnosis would therefore be entirely relevant.
Aggressiveness of prostate cancer: this is assessed by the Gleason score, which is determined by microscopic analysis of prostate biopsies. Establishing a Gleason score requires analyzing many slides and again takes time. Studies have shown a good agreement between the evaluation made by a pathologist and that of an artificial neural network.

…even a true colleague

In addition to its help with repetitive tasks where human expertise contributes little, AI has specific advantages in terms of the amount of information it can process. It is thus able to extract additional data relevant to patient care, which are certainly available routinely but often "hidden" because they are undetectable to the human eye.

AI is effective in counting tumor cells, especially in division (as here). It could also associate microscopic aspects and specific genetic mutations of cancer.
Al-Janabi S et al., CC BY

The best-known examples are the identification of genetic or genomic abnormalities in cancers, and the further evaluation of prognosis and response to treatment.

A diagnosis of cancer is usually made from the analysis of a tumor (after its biopsy or excision), placed on glass slides for study under the microscope, as we noted above. Already rich in information, these first examinations can be supplemented by genetic analyses: by identifying specific mutations of the tumour, they make it possible to better characterize it. Specialists are thus better able to set up an adequate treatment. But these additional analyzes “consume” tumor tissue and take time.

The mere observation of the digitized slides could allow algorithms to detect the relevant mutations, without resorting to genetic analysis. This saves time, money and tumor material (“tissue saving”) – the latter can be saved for other analyses.

The detection of mutations is possible by correlating a tumor shape or architecture (seen under a microscope) with the presence of mutations previously identified by DNA sequencing (reading). The algorithm must learn to associate microscopic aspects and mutations.

The same learning could be implemented to link microscopic aspects and drug response or prognosis.

Limits still strong

Even if AI will certainly improve in the medium term the diagnosis of cancers and the care of patients, the development of adequate algorithms is long and costly.

Many examples of images (ideally several thousand), normal and pathological, are indeed necessary in order to constitute the different sets on which it will be trained. This requires large databases, where each example has been annotated by a pathologist – and these image collections require large storage capacities and their digitization-annotation represents a substantial budget.

The performance of the AI depends on the quality of the data provided during its training, which makes it not free from bias. It can even amplify biases present in the training sets. And, like a well-trained human eye, it can make mistakes.

Finally, the future implementation of these digital models alongside doctors in the "real" care of patients will require the definition of standards and a legal framework, as was the case for genetic analyzes following the advent of high-throughput sequencing.

In fact, this development will require the sharing of certain medical data, which comes up against ethics and medical secrecy. Their sharing between centers is necessary for the establishment of large databases, themselves necessary for the development of reliable algorithms. And if the data is always anonymized, its possible transfer by Cloud poses confidentiality problems (risk of hacking).

Furthermore, to allow real-time assessment of disease prognosis and treatment response, the algorithms should be able to operate directly from the electronic medical record. This can only be done by respecting the recommendations of the European Medicines Agency which have yet to be established.

Future prospects

Despite these obstacles, the transition has begun. Ultimately, the goal is for AI integrates multimodal data, from the four strata of modern oncology: microscopy, radiology, genetics and clinical practice. This integration will lead to more efficient models, in particular for the evaluation of the prognosis. Within five years, AI could leave the field of research and be used in routine care.

The advent of digital pathology promises to be, in any case, a major turning point for the benefit of patients.

Audrey Rousseau, Professor in Pathological Anatomy - Physician teacher-researcher at the University Hospital of Angers, University of Angers et Leslie Tessier, PhD student, intern in pathological anatomy and cytology, RadboudUMC, Nijmegen, University of Angers

This article is republished from The Conversation under Creative Commons license. Read theoriginal article.