Breast calcification detection based on multichannel radiofrequency signals via a unified deep learning framework

Abstract

Breast calcifications in radiographic images suggest a high likelihood of breast lesion malignancy. However, it is difficult to detect calcifications in traditional B-mode ultrasound images due to resolution limits and speckle noise. In this paper, we propose a unified deep learning framework for automatic calcification detection based on multichannel ultrasound radio frequency (RF) signals. First, beamforming is used during preprocessing to merge and blend multichannel signals into one-channel RF signals. Each scan line is converted into a spectrogram by the short-time Fourier transform (STFT) to utilize the frequency domain characteristics. Then, an improved fully convolutional neural network called the RF signal Spectrogram-Calcification-Detection-Net (SCD-Net) is proposed to detect calcifications from spectrograms. This method employs a deep learning architecture based on YOLOv3 and combines features via convolutional long short-term memory (ConvLSTM). Next, a Kalman filter for tracking calcifications between consecutive spectrograms based on SCD-Net detection results is applied since the spatial coherence of calcifications in neighboring frames can be taken into account. Finally, the detected calcification is mapped from the time domain of spectrograms to B-mode images for clinical diagnosis. Experiments were conducted on a database of 337 experienced doctor-marked breast tumors with calcifications. Compared to the state-of-the-art methods for detecting calcifications, the proposed method achieved an average precision (AP) of 88.25%, an accuracy of 84% and an F1 score of 91%. The experimental results demonstrate that the unified framework has great performance for tumor calcification detection. The system can be effectively applied in a portable ultrasound instrument to accurately help radiologists and provide guidance for breast tumor diagnosis. This implies that the proposed approach can be implemented in real practice for analyzing breast RF signals, which have many useful medical applications in clinical breast tumor diagnosis.