AXTI Lab

9/ Read the full article in @Radiology_RSNA at: 📖 pubs.rsna.org/doi/epdf/10.11… #Radiology #Tweetorial #BreastImaging

8/ 🧠The authors suggest regular review sessions of AI-flagged cases to help radiologists build intuition around AI decisions. Trust in AI comes from understanding.

7/ 👀 Interesting twist: Unlike automation bias, where humans tend to over-trust AI, this study found that the consensus panel (which had access to both human & AI opinions) was more likely to disagree with the AI than the radiologist.

6/ 💡So what does this mean? Cases flagged by radiologists were more likely to be recalled than those flagged by AI. But… the AI-flagged recalls had a higher PPV. #AIvsHuman

5/ 📈 Key results: --1 radiologist: recall rate = 14.2%, PPV = 3.4% --AI only: recall rate = 4.6%, PPV = 22% --2 radiologists: recall rate = 57.2%, PPV = 2.5% --1 radiologist + AI: recall rate = 38.6%, PPV = 25% --2 radiologists + AI: recall rate = 82.6%, PPV = 34.2%

4/ 📌The authors did a secondary analysis comparing consensus outcomes of mammograms flagged by one reader (radiologist or AI), two readers (two radiologists or one radiologist + AI), and three readers. Key question: What happens at consensus discussion for those flagged cases?

3/ 🔙 In 2023 the ScreenTrustCADtrial published its promising results, compared with standard-of-care radiologist double reading, replacing one radiologist with AI for independent reading resulted in a 4% higher noninferior cancer detection.

2/ 📊 In this prospective study of 54,991 women, screening mammograms were read independently by two radiologists and one AI CAD reader. If at least one reader flagged the examination it went to a consensus panel for recall decisions.

1/ New insights from the ScreenTrustCAD trial! How do radiologists interact with AI in breast cancer screening and what are the consequences for recall rates and positive predictive values (PPV)? Let's dive in. 👇 #Radiology #AI #breastcancer