Who is best at recognizing a tumor in an image: a computer or a human? The contrast between artificial intelligence (AI) and human intelligence is a complex issue, but this is perhaps not the right question to start from. Because, taking into account that the first learns from the second, we are rather working on models that establish an alliance between the two. And we probably won’t have to wait too many years before seeing them work together reading mammograms for screening, as we tell you this week in the Breast Health newsletter.
Breast cancer, a blood draw reveals how it will change
by Tiziana Moriconi
A question of trust
AI systems capable of reading medical images – radiographs, ultrasound scans, magnetic resonances, CT scans – represent a large field of application and, therefore, a large market. Numerous diagnostic algorithms have been developed and over a hundred are already approved for clinical use by the US regulatory agency (FDA). Diagnostic software that has obtained CE certification is also available in Europe. However, there are still important issues to be addressed.
Some AI systems, for example, may turn out to be less “good” (ie not having optimal performance) when they are tested on images of hospitals other than those in which they have been “trained”, with different caseloads and equipment. Furthermore, in many cases, the algorithm proposes a decision without explaining the reasons, and it becomes difficult for clinicians to trust in a poke. This is the problem of the so-called black box.
A new therapy for patients with Brca mutations
But it’s not a race between man and machine. The challenge is how to make these systems truly useful and usable in clinical practice. And the right question becomes: how can algorithms win the trust of humans to the point of being considered as a colleague with whom to confront in doubtful cases?
For the scientists who analyze the data and the breast radiologists at Duke University (and beyond), the answer lies in making transparent and clear the decision-making steps that lead the “machine” to determine whether what is showing the radiological image is a tumor or no, and the degree of uncertainty.
The platform they have developed, and which they describe on Nature Machine Intelligence, analyzes suspicious lesions on mammography and indicates whether or not a biopsy is needed. She is certainly not the first to do so, but her algorithm shows radiologists exactly how she comes to conclusions.
Breast cancer: how to find out in advance
by Tiziana Moriconi
Duke University’s algorithm was developed by asking radiologists to label images for learning, replicating the method they themselves use to distinguish lesions in clinical practice. Their AI system worked on a database of over 1,100 mammograms of nearly 500 women (a good number, though not exceptionally high), focusing specifically on the margins of the nodules.
The point, however, is not this, but the way this platform communicates with doctors: it says why a lesion is suspected (or not suspected) and on which previous cases it relied on to establish it. Numerous software on the market – say the authors of the study – is instead of the black box type: this is the result, period.
Breast cancer: chemo, radius and acupuncture. Tuscany brings complementary therapies into the treatment path
by Tiziana Moriconi
More sophisticated software for reading ultrasound scans
“This study addresses a methodological aspect that we call in jargon explainability, and is part of a broad debate, “he comments Francesco Sardanelli, Professor of Radiology at the University of Milan and director of the Radiology Unit of the IRCCS Policlinico San Donato in Milan. In collaboration with Isabella Castiglioni, Professor of Medical Physics at the University of Milan-Bicocca and at the startup Deep Trace Technologies created by the University Institute of Higher Studies of Pavia, a software of machine learning for the evaluation of ultrasound images of breast lesions.
This AI software, which considered about 900 ultrasound images of breast lesions, is also explainable: connects the “radiomic” characteristics of the images detected by the software to the description and standardized diagnostic categories used by the breast radiologists. The study, just published on Diagnostic, showed the software’s potential to reduce unnecessary biopsies by 15-18%. Nationwide, that would be several thousand fewer biopsies every year. “Avoiding unnecessary examinations, especially biopsies, is one of the goals of AI systems in diagnostic radiology,” explains Sardanelli.
“To achieve this result on a large scale, a close collaboration between clinicians and data scientists will be required. More generally, we will have to aim at improving the quality of the radiological work. The reading of the images will be supported by AI systems, with human intelligence that he must maintain a guiding role, of last decision and of responsibility “.
Objective: to improve the service to patients
In short, the report will always be signed by a human being, who in addition to experience makes his professional ethics and empathy available, fundamental aspects when communicating the diagnosis and, above all, the uncertainty of a diagnosis. “The reading time saved – continues the expert – can allow more time to speak with patients, communicate the need for further information, discuss cases in multidisciplinary meetings, increase screening coverage”. The process to make what has just been described a reality therefore depends above all on health policies.
“I hope – concludes Sardanelli – that with regard to the use of AI in the reading of mammography screening, a concertation of regional screening systems can be made and therefore at least a national, if not European, address can be had”.