Weakly supervised multi-needle detection in 3D ultrasound images with bidirectional convolutional sparse coding

Accurate and automatic multi-needle detection in three-dimensional (3D) ultrasound (US) is a key step of treatment planning for US-guided prostate high dose rate (HDR) brachytherapy. In this paper, we propose a workflow for multineedle detection in 3D ultrasound (US) images with corresponding CT images used for supervision. Since the CT images do not exactly match US images, we propose a novel sparse model, dubbed Bidirectional Convolutional Sparse Coding (BiCSC), to tackle this weakly supervised problem. BiCSC aims to extract the latent features from US and CT and then formulate a relationship between them where the learned features from US yield to the features from CT. Resultant images allow for clear visualization of the needle while reducing image noise and artifacts. On the reconstructed US images, a clustering algorithm is employed to find the cluster centers which correspond to the true needle position. Finally, the random sample consensus algorithm (RANSAC) is used to model a needle per ROI. Experiments are conducted on prostate image datasets from 10 patients. Visualization and quantitative results show the efficacy of our proposed workflow. This learning-based technique could provide accurate needle detection for US-guided high-dose-rate prostate brachytherapy, and further enhance the clinical workflow for prostate HDR brachytherapy.