Human‒machine interaction based on real-time explainable deep learning for higher accurate grading of carotid stenosis from transverse B-mode scan videos

Abstract

Objectives

We aim to develop an explainable deep learning (DL) model to assist radiologists in carotid stenosis classification by providing understandable or explainable output.

Materials and methods

This prospective study included patients suspected ≥50 % carotid stenosis from three hospitals between February 2022 and October 2022. The DL model CaroNet-Dynamic 2.0 was trained based on carotid transverse ultrasound (US) videos. Model performance was evaluated using expert (with 15 years of experience in carotid US evaluation) diagnoses as the reference standard. Finally, CaroNet-Dynamic 2.0 was integrated into a user-friendly web graphical user interface to support artificial intelligence (AI) explainability and human supervision. The human‒machine interaction strategy was evaluated with five senior and five junior radiologists. Area under the receiver operating characteristic curve (AUROC) were calculated.

Results

A total of 311 patients (mean age ± standard deviation, 71.3 years ± 8.3; 247 men) were included. CaroNet-Dynamic 2.0 showed robust performance in carotid stenosis classification and approached that of senior radiologists (P > 0.05 for all comparisons). Junior and senior radiologists initially disagreed with AI on 47 and 37 plaques, respectively. Using the human‒machine interaction, they adopted AI diagnoses for 38 and 28 plaques, overruling 9 each. The AUROCs of human‒machine interaction achieved 0.868–0.896 and 0.875–0.904 for junior and senior radiologists respectively, substantially outperforming junior radiologists alone (P < 0.05 for all comparisons).

Conclusion

CaroNet-Dynamic 2.0 attempted to explain to radiologists the information the DL model used to make decisions and proactively involved them in the decision loop to further improve their performance.