Monday, November 30 | 9:30 a.m.-10:30 a.m. | M3-SSGI03-2 | Room E451B
In this scientific presentation, researchers will describe how radiologists and AI can make an effective team for detecting pancreatic cancer on CT.
Presenter Tijmen De Haas of Radboudumc in the Netherlands, and colleagues set out to identify where radiologists and AI systems have trouble detecting pancreatic cancer on contrast-enhanced CT, and whether their strengths complement each other.
Using data from the international PANORAMA reader study, the group compared how 68 radiologists and AI software performed on 391 CT scans, 144 of which showed pancreatic cancer. The radiologists also rated how difficult they found each case.
AI localized lesions more consistently than the radiologists, and its advantage grew as cases became harder. On cases rated very challenging, it made 13% fewer errors than radiologists. Radiologists, however, held a slight edge on cases rated not challenging and performed even better on cases where the AI was uncertain. Radiologists who read a higher volume of cases also performed better, while years of experience made no difference.
“Combined with the association between lower case volume and weaker radiologist performance, these findings support a framework in which AI confidence, perceived case difficulty, AI-based lesion localization, and radiologist case volume may guide effective combined reading strategies,” the researchers wrote. “Understanding where AI and radiologists succeed or fail may help optimize clinically safe integration of AI into pancreatic cancer detection workflows.”
Find out more when this research is presented on Monday morning.
