Multicenter DL Model Improves Radiologists' Precision for Subcentimeter Breast Cancer on MRI

In a multicenter reader study, adding deep learning to breast MRI interpretation raised radiologists' precision for subcentimeter (≤1 cm) invasive cancers from 72.9% to 83.2% without a significant change in sensitivity (89.5% vs 86.8%, p>0.05).

• The stand-alone deep learning model for ≤2 cm cancers had an AUPRC of 0.42, sensitivity 83.2%, and precision 33.2%.
• For subcentimeter lesions, radiologists outperformed the model (F1 score 0.80 vs 0.33).
• Limitation: Readers were informed each exam contained a single cancer, and the cancer-enriched dataset may limit generalizability.

Automated summary

RadPigeon summaries are original and for information only. They are not clinical advice.