Sequential RF-based curriculum learning boosts breast ultrasound detection in small-data settings

A Phase-Envelope-B-mode curriculum transfer learning approach using RF-derived representations improved breast ultrasound tumor detection by 2.08% relative mean average precision compared to B-mode training alone, while maintaining a single-model inference pathway in a controlle…

• The Phase-Envelope-B-mode (P-E-B) curriculum strategy achieved a 2.08% relative improvement in mAP@50 over single B-mode training under a fixed patient-level split with 100-seed repeated evaluation.
• The approach accelerates convergence without requiring multimodal input during inference, using only standard B-mode images.
• Results are proof-of-concept within a controlled small-data setting; external clinical generalizability was not assessed.

Automated summary

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