A group of scientists from Tianjin Medical University (Tmu), in China, has just developed a digital tool, based on an artificial intelligence algorithm, capable of analyzing images of metastatic cells and identifying the location of the primary tumor with greater precision than that of human oncologists. A model, write the authors of the research on pages of the magazine Nature Medicinewhich could help improve the diagnosis and treatment of cancers in more advanced stages, and potentially increase patients’ life expectancy.
Targeted screening and early diagnosis: this is how AI is changing medicine by Sandro Iannaccone 02 February 2024
Where does cancer come from?
Knowledge of the origin of a certain metastasis is a crucial factor in increasing the probability of success of therapies. Unfortunately, many tumors are able to grow without giving any symptoms and without being detected, and spread to organs and regions of the body even very far from their origin: for this reason, in some cases it is very difficult to reconstruct the “path” of the metastasis and trace back to the primary tumor. In approximately five cases out of a hundred, the origin of the tumor cannot be identified, and the prognosis for the patient is almost always unfavorable. At the moment, the most effective diagnostic method for “unmasking” metastases is based on a similarity search: the cells of a lung metastasis from a breast tumor, for example, “resemble” breast tumor cells in some way . It is by analyzing these similarities, precisely, that oncologists try to infer the location of the primary tumor.
Artificial intelligence, here’s what can be done in healthcare by Gianluigi Grieco, Piero Poccianti 15 March 2024
The gaze of the Ai
This is where artificial intelligence comes into the picture, which – among other things – has become increasingly proficient in the field of image recognition and classification in recent years: Tian Li Xiangchun and colleagues, a group of Tmu bioinformaticians specializing in deep learning, have developed an algorithm specialized in the analysis of images of metastatic cells. The model was trained on approximately 30 thousand images of cells taken from the lung or abdominal fluid of over 21 thousand cancer patients in which the origin of the primary tumor was known: in this way the algorithm “learned” to associate a specific image with a specific origin; subsequently, it was tested on another 27 thousand images – always relating to patients with a known primary tumor – and was able to correctly identify the tumor source in 83% of cases. The model, in truth, provides a list of possible “candidates” for the primary tumor, ordered by probability: in 99% of cases the correct answer was found to be in the first three answers of the algorithm.
The wearable AI that restores the voice in case of vocal cord problems by Simone Valesini 21 March 2024
A list of possibilities
Have a list of possibilities for the primary tumor, they explain above Nature, is very useful for oncologists, because it reduces the number of additional – and often invasive – tests necessary for the correct diagnosis. In the newly published study, predictions were limited to 12 cancer types (including lung, ovarian, stomach and breast) because other types (including prostate and kidney) usually do not spread into lung and abdominal fluids; but one could think, in the future, of extending the analysis to cells taken elsewhere, in order to broaden the diagnostic possibilities of the system, as well as integrating it with instruments that work with similar principles but analyze other information (for example tissue samples or genomic data) to further improve its accuracy.