What will be the place of artificial intelligence (AI) in medicine tomorrow? Diagnostic problems are already Dr. Could it be solved not by House, but by his digital alter ego? In fact, AI is already outperforming the eye in detecting certain tumors from radiological images (mammograms, MRIs)… This has led some to say that computers will soon replace human experts.
However, contrary to these predictions, the radiologist did not disappear: on the contrary, an unexpected “collaboration” arose between him and the machine that would take his place. The former works to channel the capabilities and strengths of the latter to improve interpretation and diagnosis for the benefit of patients.
This question, which helps to make a correct diagnosis, is central and is applied both in psychiatry, where AI is taking its first steps, and in oncology… In pathological anatomy, that is, “the examination of organs, tissues, or cells to identify and analyze their association with disease (cancer and etc.) related abnormalities”, prospects and promises are great.
Is artificial intelligence already capable of such analysis? Can a human be more efficient than an expert?
Misunderstandings and confusion abound, and it’s important to understand why. We offer you exactly this point here.
What enabled the first steps of “digital pathology”.
Like any human expert, the diagnosis for AI is based, among other things, on an object as simple as it is important: glass slides on which the pathologist places a very thin “slice” of the tissue to be analyzed (lung, liver, etc.), for observation under a microscope. .
Through this microscopic analysis, a pathologist can identify different cell types, compare their shape, or even their spatial structure (architecture) to identify abnormal masses—for example, tumors.
The mass digitization of these slides paved the way for the use of AI in pathological anatomy. The advent of adaptive scanners makes it possible to acquire and store microscope slides in digital form in a growing number of hospitals. However, the original slides are preserved…due to storage costs, this will not be possible for all digital versions of them.
This procedure, which led to “digital pathology”, made it possible to work on algorithms designed to perform their analysis in an automated way. With the aim of AI being able to assist the pathologist in his diagnosis. It is also useful for ergonomic reasons and to save time.
But like a human, a machine (mostly artificial neural networks) needs to be trained. First, he must be able to “look” at the blades and understand what it is. This analysis uses pattern recognition technology as a core technique.
Second, he must be able to interpret what he “sees”. Artificial intelligence is based on the concept of learning and the ability to make inferences, that is, to transfer the knowledge gained during its formation and training to other situations that are comparable, but not similar: for example, to recognize breast cancer lymph node micrometastasis (a cluster of several tumors) to the eye, having previously seen other images of metastases. cells that do not touch.
It should be noted that digital slides contain many more pixels than radiological images and contain thousands of cells – so they are particularly rich in information that algorithms can use.
[Près de 80 000 lecteurs font confiance à la newsletter de The Conversation pour mieux comprendre les grands enjeux du monde. Abonnez-vous aujourd’hui]
A fast and reliable digital assistant…
Current research and testing suggests that artificial intelligence may ultimately be relevant in several areas:
- automation of the most repetitive and subjective activities,
- helps in tumor detection, aggressiveness assessment and subtyping,
- counting tumor cells, especially those that divide (mitoses);
- assessment of the intensity of the immune response (the number of lymphocytes attacking the tumor).
The interests are multifaceted: giving the human pathologist time to devote himself to the most complex cases where human added value is real, making the final diagnosis faster and more reliable. And most importantly, the results of AI analyzes in science are generally reproducible.
We can already identify specific cases where the contribution of AI is relevant:
- Breast cancer detection: Algorithms outperform pathologists in detecting micrometastases in axillary lymph nodes.
- Assessment of breast cancer prognosis: artificial neural networks effectively identify cell markers made using specific antibodies (immunohistochemistry technique). In breast cancer, determining the amount of HER2 protein expression in tumor cells allows to evaluate the prognosis of the disease and the response to certain drugs – this protein stimulates the development of cancer. Therefore, computer-aided diagnosis would be absolutely relevant.
- Aggressiveness of prostate cancer: assessed by Gleason score, determined by microscopic analysis of prostate biopsies. Generating a Gleason score requires analyzing many slides and is again time-consuming. Studies have shown good agreement between a pathologist’s assessment and an artificial neural network assessment.
…even my real colleague
In addition to helping with repetitive tasks where human expertise contributes little, AI has particular advantages in terms of the amount of information it can process. In this way, it is able to extract additional information relevant to patient care that is certainly routinely present, but is often “hidden” because it cannot be detected by the human eye.
The most popular examples are the identification of genetic or genomic abnormalities in cancers and the subsequent assessment of prognosis and response to treatment.
The diagnosis of cancer is usually made from the analysis (after biopsy or excision) of a tumor placed on glass slides for examination under a microscope as mentioned above. These initial examinations, already rich in information, can be complemented by genetic analyses: by identifying specific mutations of the tumor, it allows to better characterize it. Specialists are thus better able to establish adequate treatment. But these additional analyzes “consume” the tumor tissue and take time.
Mere observation of digitized slides may allow algorithms to detect relevant mutations without resorting to genetic analysis. This saves time, money and tumor material (“tissue sparing”) – the latter of which can be saved for other analyses.
Detection of mutations is possible by correlating tumor shape or architecture (seen under a microscope) with the presence of mutations previously identified by DNA sequencing (arrow). The algorithm must learn to correlate microscopic aspects and mutations.
The same study can be done to correlate microscopic aspects and response to medication or prognosis.
The limits are still strong
Even if AI improves cancer diagnosis and patient care in the medium term, developing adequate algorithms is time-consuming and expensive.
Many examples of normal and pathological images (ideally several thousand) are really needed to organize the different sets on which it will be trained. This requires a large database where each specimen is annotated by a pathologist – and these image collections require large storage capacities and their digitization-annotation represents a significant budget.
The performance of AI depends on the quality of the data provided during training, which does not make it free from bias. It can even reinforce biases present in the training sets. And like the well-trained human eye, it can make mistakes.
Finally, the future application of these digital models in the “real” care of patients alongside physicians will require the establishment of standards and a legal framework, as is the case for genetic analysis following the advent of high-throughput sequencing.
In fact, this development would require the sharing of certain medical information, which is contrary to ethics and medical confidentiality. Their distribution among centers is necessary for the creation of large databases necessary for the development of reliable algorithms. If the data is always anonymized, its possible transmission through the Cloud creates privacy problems (risk of hacking).
In addition, algorithms must be able to work directly from the electronic medical record to allow real-time assessment of disease prognosis and treatment response. This can only be done by following the yet to be established recommendations of the European Medicines Agency.
Despite these obstacles, the transition has begun. Ultimately, the goal is to integrate multimodal data from the four layers of modern oncology for AI: microscopy, radiology, genetics, and clinical practice. This integration will lead to more efficient models, especially for forecasting. Within five years, artificial intelligence may leave the research field and be used in everyday care.
The advent of digital pathology, however, promises to be a major turning point for patients.
Audrey RussoProfessor of Pathological Anatomy – doctor-teacher-researcher at Angers University Hospital, University of Angers and Leslie TessierPhD student, intern in pathological anatomy and cytology, RadboudUMC, Nijmegen, University of Angers
This article is republished from The Conversation under a Creative Commons license. Read the original article.
SUPPORT INFORMATION CHRÉTIENNE
As Info Chrétienne is an online press service recognized by the Ministry of Culture, your donation is up to 66% tax deductible.