5 Key Takeaways

• 1

  An interpretable AI model may match the diagnostic accuracy of standard deep-learning methods for breast lesion classification.

• 2

  The study evaluated a concept bottleneck model using data from 1,695 lesions across five hospitals from 2016 to 2025.

• 3

  The CBM achieved an AUC of 0.92 in the test set and maintained high performance with an AUC of 0.93 in external validation.

• 4

  Radiologist accuracy improved significantly with CBM assistance, increasing from 71% to 91% and reducing interpretation time.

• 5

  The model demonstrated clinical utility by aiding in the correct downgrading of 22% of suspicious benign lesions.