A new “combined” Artificial Intelligence (AI) seems to be able to very accurately detect the presence of thyroid cancer and to make predictions on the characteristics of the disease both from a pathological and a genomic point of view. All by analyzing routine ultrasound images: a method that promises to carry out screening, tumor staging and planning of personalized treatments at low cost and in a non-invasive way.
A combined AI system
As illustrated during the multidisciplinary Symposium on Head and Neck Tumors by the team from Massachusetts General Hospital and Harvard Medical School that developed it, the new AI brings together different methods of machine learning: radiomics, which allows to extrapolate information quantitative as shape and volume from medical images; topological data analysis (Tda); a deep learning system based on algorithms that analyze data on multiple levels of a neural network; a Ti-Rads pattern analysis algorithm (a classification for describing thyroid nodules during ultrasound examination).
Machine learning: 4 is better than 1
The system was trained on hundreds of ultrasound images of suspected thyroid nodules, both malignant and benign. Another set of images was then used to validate its accuracy, obtained by comparing the results returned by the AI and those of the diagnoses with traditional methods (biopsies, surgical reports, genomic sequencing).
The four methods combined accurately diagnosed 98.7% of the malignant thyroid nodules used for internal testing, surpassing the results achieved by each method used individually. In particular, radiomics alone is able to correctly identify 89% of malignant tumors, deep learning 87%, Tda 81% and machine learning based on Ti-Rads 80%. Accuracy tested on external image sets drops slightly (93%).
The accuracy in predicting the stage of the disease according to the Tnm classification is also high (93% for the T stage, 89% for the N stage and 98% for the M stage).
“We have developed an artificial intelligence platform that examines ultrasound images and predicts with high accuracy whether a potentially problematic thyroid nodule is cancerous,” explains Annie Chan of Mass General Hospital. “If it’s cancerous, we can further predict the stage of the cancer, lymph node involvement, and the presence or absence of a BRAF mutation.”