Artificial Intelligence (AI) TI-RADS

AI TI-RADS is a data driven analysis and revision of the 2017 ACR TI-RADS ¹. Published in May 2019 ², this had the intention of simplyfing categorisation and improving specificity while maintaining high sensitivity. This system used a training set of 1325 nodules with known cytology and a genetic learning algorithm to optimise the current lexicon and scoring structure. This system was then applied by expert and nonexpert radiologists and compared with known ACR TI-RADS categories.

Changes to ACR TI-RADS

The five imaging characteristic categories were maintained with the following changes in each:

• composition - only solid nodules receive points
  • cystic and spongiform nodules again receive 0 points, but AI TI-RADS found this sufficient to assign benignity regardless of findings in any of the other categories
  • mixed cystic / solid nodules received 0 points instead of 1 point
  • solid or near completely solid nodules received 3 points instead of 2 points
  • can't classify 0 points instead of 2 points
• echogenicity - only hypoechoic nodules receive points
  • iso / hyperechoic and can't classify nodules 0 points instead of 1 point
• shape
  • taller-than-wide nodules only 1 point instead of 3 points
• margin
  • no change
• echogenic foci
  • macrocalcification receive 0 points instead of 1 point

The point level for each TR category and recommendation for FNA by size remained the same.

Outcome comparison with ACR TI-RADS

Tested against 100 nodules, sensitivity was identical between AI and ACR TI-RADS (93%). Specificity was increased for AI TI-RADS at 65% compared with 47% with a single expert reader, and also an increase of 55% from 47% in a nonexpert reader group.

43 nodules were down categorised, with 15 nodules not meeting requirement for FNA (all of which were benign). No extra malignancies were missed using the AI TI-RADS scoring algorithm.

See also

• ACR TI-RADS