Artificial Intelligence (or AI) is playing an ever increasing role in medicine, especially in the oncology sector, as reported recently in three of the leading worldwide scientific journals: Science, which dedicates an in-depth article to the subject, as well as Nature and JAMA Network Open, which focus on the revolution taking place in the interpretation of mammograms.
Nature, in particular, published the results of a study conducted by Google Health experts (in Palo Alto, California), in which a large number of mammograms taken in Great Britain and in the United States were reinterpreted by artificial intelligence systems. The result was very promising: the programs they used, in fact, reduced the number of false positives in the USA by 5.7% and those in Great Britain by 1.2%, and the false negatives respectively by 9.4% and 2.7% (false positives report the presence of a tumor, even if the cells, in reality, are not really cancerous, whereas false negatives, do the opposite). Furthermore, in a comparison of the interpretation of the mammograms by six radiologists and by artificial intelligence systems, the latter won again. Finally, in a simulation of the British system, which uses a double reading system (i.e. where the mammogram is read by two radiologists, one of which, in this case, was replaced by AI), the non-inferiority of the computerised system compared to the human eye was confirmed and showed that artificial intelligence can reduce the workload of the person doing the second reading by 88%.
In the study published by JAMA Network Open the situation was less skewed towards AI, but still pointed in the same direction. This time, researchers from an international consortium compared human readings with those of various types of algorithms, within the context of an international “competition” called Digital Mammography DREAM Challenge, launched by IBM Research, Sage Bionetworks and Kaiser Permanente Washington Health Research Institute. To this end, more than 144,000 mammograms from 85,000 US women and 166,000 mammograms from 68,000 Swedish women were examined by various types of algorithms, with the contribution of more than 1100 researchers spread over 126 groups in 44 countries, who offered to apply various protocols to the examinations already carried out (for example, by combining or using single previous clinical examinations or mammograms of the same person).
In the end, none of the algorithms put to the test was able to improve the performance of the radiologists alone. However, the integration of the two types of reading – one human and one computerized – resulted in a reduction of the number of false positives. Therefore, beyond the technical details and the type of algorithm, what definitely appears to be making headway is the increasingly important role played by systems that help to interpret radiological examinations, starting with mammograms, thus offering a potentially, significant advantage for all the women that participate in screening programs.
“Our study”, says Gustavo Stolovitzky, founder of DREAM Challenge “suggests that an algorithmic combination of AI and radiologist interpretations could provide a mechanism for significantly reducing unnecessary diagnostic workups for at least half a million women in the United States alone”.