Tag: AI generalizability

CXR AI’s Screening Generalizability Gap

Posted on by Jake Fishman

A new European Radiology study detailed a commercial CXR AI tool’s challenges when used for screening patients with low disease prevalence, bringing more attention to the mismatch between how some AI tools are trained and how they’re applied in the real world.

The researchers used an unnamed commercial AI tool to detect abnormalities in 3k screening CXRs sourced from two healthcare centers (2.2% w/ clinically significant lesions), and had four radiology residents read the same CXRs with and without AI assistance, finding that the AI:

  • Produced a far lower AUROC than in its other studies (0.648 vs. 0.77–0.99)
  • Achieved 94.2% specificity, but just 35.3% sensitivity
  • Detected 12 of 41 pneumonia, 3 of 5 tuberculosis, and 9 of 22 tumors 
  • Only “modestly” improved the residents’ AUROCs (0.571–0.688 vs. 0.534–0.676)
  • Added 2.96 to 10.27 seconds to the residents’ average CXR reading times

The researchers attributed the AI tool’s “poorer than expected” performance to differences between the data used in its initial training and validation (high disease prevalence) and the study’s clinical setting (high-volume, low-prevalence, screening).

  • More notably, the authors pointed to these results as evidence that many commercial AI products “may not directly translate to real-world practice,” urging providers facing this kind of training mismatch to retrain their AI or change their thresholds, and calling for more rigorous AI testing and trials.

These results also inspired lively online discussions. Some commenters cited the study as proof of the problems caused by training AI with augmented datasets, while others contended that the AI tool’s AUROC still rivaled the residents and its “decent” specificity is promising for screening use.

The Takeaway

We cover plenty of studies about AI generalizability, but most have explored bias due to patient geography and demographics, rather than disease prevalence mismatches. Even if AI vendors and researchers are already aware of this issue, AI users and study authors might not be, placing more emphasis on how vendors position their AI products for different use cases (or how they train it).

The Mammography AI Generalizability Gap

Posted on by Jake Fishman

The “radiologists with AI beat radiologists without AI” trend might have achieved mainstream status in Spring 2020, when the DM DREAM Challenge developed an ensemble of mammography AI solutions that allowed radiologists to outperform rads who weren’t using AI.

The DM DREAM Challenge had plenty of credibility. It was produced by a team of respected experts, combined eight top-performing AI models, and used massive training and validation datasets (144k & 166k exams) from geographically distant regions (Washington state, USA & Stockholm, Sweden).

However, a new external validation study highlighted one problem that many weren’t thinking about back then. Ethnic diversity can have a major impact on AI performance, and the majority of women in the two datasets were White.

The new study used an ensemble of 11 mammography AI models from the DREAM study (the Challenge Ensemble Model; CEM) to analyze 37k mammography exams from UCLA’s diverse screening program, finding that:

  • The CEM model’s UCLA performance declined from the previous Washington and Sweden validations (AUROCs: 0.85 vs. 0.90 & 0.92)
  • The CEM model improved when combined with UCLA radiologist assessments, but still fell short of the Sweden AI+rads validation (AUROCs: 0.935 vs. 0.942)
  • The CEM + radiologists model also achieved slightly lower sensitivity (0.813 vs. 0.826) and specificity (0.925 vs. 0.930) than UCLA rads without AI 
  • The CEM + radiologists method performed particularly poorly with Hispanic women and women with a history of breast cancer

The Takeaway

Although generalization challenges and the importance of data diversity are everyday AI topics in late 2022, this follow-up study highlights how big of a challenge they can be (regardless of training size, ensemble approach, or validation track record), and underscores the need for local validation and fine-tuning before clinical adoption. 

It also underscores how much we’ve learned in the last three years, as neither the 2020 DREAM study’s limitations statement nor critical follow-up editorials mentioned data diversity among the study’s potential challenges.

Get every issue of The Imaging Wire, delivered right to your inbox.

You might also like..

Select All

You're signed up!

It's great to have you as a reader. Check your inbox for a welcome email.

-- The Imaging Wire team

You're all set!