Abstract 734P
Background
Ultrasound-based artificial intelligence (AI) model provides a promising tool for cancer diagnosis, but is hardly generalizable to the low-quality image captured by inexperience sonographer, hindering the adoption of AI in point-of-care setting and low-income countries.
Methods
We conducted a diagnostic study (CAPTURE) involving 3814 adult patients from 5 hospitals who had thyroid nodule detected during 2018-2020. A baseline ultrasonographical deep learning model was developed based on previously available dataset. A novel AI model which automatically integrates multi-images from the same subject is developed to overcome the inexperience of sonographers. The diagnostic performance of AI models was evaluated using area under the curve (AUC) and compared among image subgroups from experienced and inexperienced sonographers.
Results
In our validation cohort, the AI model has significantly lower AUC in images captured by inexperienced sonographer compared to experienced ones (AUC=0.70 for <3 years vs 0.89 for ≥3 years, p=0.013). Our proposed novel model improved AI performances significantly in inexperienced sonographer subgroup from AUC 0.72 (95CI 0.64-0.79) to 0.87 (95CI 0.84-0.93, p<0.001).
Conclusions
The AI model prediction were significantly impacted by the image capture process, especially the expertise of sonographer. Automatic multi-image capturing and integration may increase the generalizability and accessibility of AI-assist diagnosis in point-of-care setting and low-income countries.
Clinical trial identification
Editorial acknowledgement
Legal entity responsible for the study
The author.
Funding
National Natural Science Foundation of China 81903417.
Disclosure
The author has declared no conflicts of interest.