Coarse-to-Fine Curriculum Transfer Learning Using RF-Derived Ultrasound Representations for Small-Data Breast Tumor...

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Deep Learning for Medical Diagnosis

Background:

• Breast ultrasound (BUS) is crucial for tumor detection but faces challenges like noise, low contrast, operator dependency, and limited datasets for deep learning.
• Raw radiofrequency (RF) signals offer richer acoustic information than B-mode images, but multimodal fusion increases computational costs.
• Existing deep learning models struggle with the inherent limitations of BUS data, necessitating innovative approaches for improved accuracy and efficiency.

Purpose of the Study:

• To propose a novel curriculum transfer learning approach for breast tumor detection using ultrasound data.
• To address the limitations of BUS data and multimodal fusion by maintaining a single detection model architecture.
• To evaluate the effectiveness of sequentially exploiting different ultrasound information representations (Phase, Envelope, B-mode) during training.

Main Methods:

• Developed a curriculum transfer learning strategy that sequentially utilizes Phase, Envelope, and B-mode images derived from raw RF signals.
• Maintained a single, lightweight detection model architecture (YOLO variants) throughout the training and evaluation process.
• Conducted extensive experiments across nine settings, including single-modality training and curriculum learning, with 100 random seeds for robust evaluation.

Main Results:

• The proposed Phase-Envelope-B-mode (P-E-B) curriculum transfer learning strategy achieved the highest average mAP@50.
• Demonstrated a 2.08% relative improvement over single B-mode training under fixed patient-level split and 100-seed evaluation.
• Observed a lower average convergence epoch compared to single B-mode training, indicating improved training efficiency.

Conclusions:

• The RF-derived representations can serve as valuable training-stage curriculum information for B-mode-based breast tumor detection.
• The proposed curriculum transfer learning approach enhances detection performance and convergence speed while maintaining a single B-mode inference pathway.
• Findings provide proof-of-concept for small-data settings, suggesting potential for improved deep learning in breast tumor detection using ultrasound.