Courtesy / Microsoft
Study presents a promising approach based on artificial intelligence. FCDD is based on the detection of anomalies rather than traditional classification.
O Cancro yes Mom It is the most common type of cancer among women around the world. There is screening that leads to early detection, and which is the most effective tool for saving lives.
Mass magnetic resonance imaging continues to be Imperfect: It has high sensitivity, but also has many false positives, increases anxiety in patients and gives rise to unnecessary biopsies.
More exams, more uncertainty and more follow -up procedures that are unnecessary. Also because only less than 5% of women who perform magnetic resonance screening are eventually diagnosed with cancer.
Artificial intelligence arrived (IA) And with her, a real opportunity to detect breast cancer earlier, reduce unnecessary interventions and even save more lives.
One new features a promising AI -based approach. The goal It is not just to detect cancer, but to do so that radiologists can trust and that patients can understand.
The Ai for Good Lab project, from Microsoft, wants to contribute to greater clarity, accuracy and confidence in the screening of breast cancer.
The model is called FCDD (Fully Convolutional Data Description) and is based on detection of anomalies instead of traditional classification.
And this is a “Significant change”, Read in a statement sent to ZAP. AND “More than a technical advance”.
Instead of trying to learn how each type of cancer is presented, the model learn how normal breast tests are and marks everything that deviates from this standard. It will be effective in real tracking contexts, where cancer is rare and the anomalies are very varied.
Three advantages
In a data set with over 9,700 mammary RM exams, the model was tested in high and low prevalence scenarios – including realistic tracking populations, where only 1.85% of the exams contained cancer.
The first advantage: greater accuracy in low prevalence populations. FCDD has overcome traditional AI models in identifying malignancy, dramatically reducing false positives. He doubled the positive predictive value of standard models and reduced the so -called “false alarms” by over 25%.
Explains, it is not limited to “yes” or “no”: generates thermal maps that visually highlight the suspicious location of the tumor in the two -dimensional projection of resonance. These explanatory maps coincided with the retrospective notes of expert radiologists with 92% accuracy.
And makes a generalization between institutions: maintained high performance in a publicly available external data set and other independent internal set; It has a strong potential for broader clinical adoption.
As a next step, the model will be tested prospectively in larger and more diverse clinical populations. However, the results are promising and represent an important change in the way the AI’s role in medicine is viewed.
