Predicting Thyroid Cancer with AI Obtains Extremely Accurate Results

A large retrospective investigation revealed that the AI-enabled ultrasound technology distinguishes benign and malignant thyroid nodules with near-perfect accuracy. The algorithm predicted malignancy with an overall accuracy of 98.7% on the combined training and validation sets. The area under the receiver operating characteristic curve (AUC), performance was 0.99 (1.00 = 100 percent accuracy).

Article image

Annie W. Chan, MD, of the Mass General Cancer Center in Boston, reported at the Multidisciplinary Head and Neck Cancer Symposium in Phoenix that the multimodal method had equal accuracy in predicting T-stage, extracapsular extension, and the presence of the BRAF mutation.

"We have created a platform to predict malignancy, surgical and genomic outcomes," MedPage Today cites Chan. "Various AI models are interacting with one another. Conducting multicenter studies with prospective validation would be ideal."

The current method of evaluating thyroid nodules is nearly entirely based on the radiologist's perception. The diagnosis is based on the object's shape, look, and size. Descriptive phrases like "concerning," "suspicious," or "worrisome" are imprecise, according to Chan.

Beyond the first evaluation, the current strategy to treat thyroid nodules is time-consuming, prone to high inter-observer variability, and invasive, as staging cannot be done using CT or MRI and instead requires ultrasound-guided biopsy.

The integrated platform incorporates many forms of image processing and artificial intelligence technology to attain more accuracy. These are the main components of the new platform:

  • Thyroid Imaging Reporting and Data System (TI-RADS)
  • Radiomics that describes how an image's texture and gray level intensities are measured.
  • TDA (Topological Data Analysis) that investigates the pixel generation kinetics.
  • Deep Learning that re-scan photos to remove noise while preserving crucial information.

Radiomics technology is well-established, and its sensitivity to picture noise is famously acknowledged. TDA provides reliable noise analysis. Deep learning does not involve image segmentation or feature selection, but its accuracy is dependent on the training set's quality.

Internal training, validation trials, and an external validation study were employed to evaluate 1,346 ultrasounds of 784 individuals with known thyroid nodule status. During the internal research, 156 malignant nodules and 357 benign neoplasms were studied, whereas 270 malignant nodules and 50 benign neoplasms were analyzed during external research.

Radiomics scored best for predicting malignancy (88.7%, AUC 0.87), followed by deep learning (87.4%, AUC 0.92), TDA (81.5 percent, AUC 0.87), and TI-RADS (80%, AUC 0.76), according to internal validation.

The integrated technological platform predicted T-stage cancers with 93 percent accuracy, extracapsular spread with 98 percent accuracy, and BRAF mutation with 96 percent accuracy. Individual system components outperformed all values.

In an external validation study, the integrated platform had a 93 percent cancer prediction accuracy and an AUC of 0.94.

According to Chan, the findings are comparable to four other AI-based thyroid cancer prediction studies published in 2020 and 2021. Three Chinese studies showed an accuracy rate of 87-96 percent, whereas one study in the United States had an accuracy rate of 82 percent.

AI models require a large amount of diverse and quality data in order to improve their accuracy and reliability. The more quality data are  employed in training of AI system, the better results can be expected. Medical Data Cloud platform specializes in producing a variety of custom medical datasets with full support from medical, IT and AI specialists. 

References:

  1. AI Platform Near Perfect for Predicting Thyroid Malignancy - by Charles Bankhead, Senior Editor, MedPage Today February 27, 2022
  2. Sánchez JF. TI-RADS classification of thyroid nodules based on a score modified according to ultrasound criteria for malignancy. Rev Argent Radiol. 2014;78(3):138-48.
  3. Paul R, Juliano A, Faquin W, Chan AW. An Artificial Intelligence Ultrasound Platform for Screening and Staging of Thyroid Cancer. International Journal of Radiation Oncology, Biology, Physics. 2022 Apr 1;112(5):e8.
Contact