Iterative Multi-domain Regularized Deep Learning for Anatomical Structure Detection and Segmentation from Ultrasound Images

Abstract

Accurate detection and segmentation of anatomical structures from ultrasound images are crucial for clinical diagnosis and biometric measurements. Although ultrasound imaging has been widely used with superiorities such as low cost and portability, the fuzzy border definition and existence of abounding artifacts pose great challenges for automatically detecting and segmenting the complex anatomical structures. In this paper, we propose a multi-domain regularized deep learning method to address this challenging problem. By leveraging the transfer learning from cross domains, the feature representations are effectively enhanced. The results are further improved by the iterative refinement. Moreover, our method is quite efficient by taking advantage of a fully convolutional network, which is formulated as an end-to-end learning framework of detection and segmentation. Extensive experimental results on a large-scale database corroborated that our method achieved a superior detection and segmentation accuracy, outperforming other methods by a significant margin and demonstrating competitive capability even compared to human performance. arXiv:1607.01855v1 [cs.CV] 7 Jul 2016Recently, fully convolutional networks (FCN), i.e, a variant of CNN, achieved the state-of-the-art performance on image segmentation related tasks [6]. Such

Artificial intelligence in radiology

“…Studies have also shown that deep learning technologies are on par with radiologists’ performance for both detection 36 and segmentation 37 tasks in ultrasonography and MRI, respectively. For the classification tasks of lymph node metastasis in PET-CT, deep learning had higher sensitivities but lower specificities than radiologists 38 .…”

Deep Learning for Cardiac Image Segmentation: A Review

“…Several approaches aim to leverage the coherence between temporally close frames to improve the accuracy and robustness of the LV segmentation. Nascimento (2010, 2013) proposed a dynamic modeling method based on a sequential monte carlo (SMC) (or particle DBN with two-step approach: localization and fine segmentation LV 2D A2C, A4C Nascimento and Carneiro (2017) deep belief networks (DBN) and sparse manifold learning for the localization step LV 2D A2C, A4C Carneiro (2014, 2019) DBN and sparse manifold learning for one-step segmentation LV 2D A2C, A4C Veni et al (2018) FCN (U-net) followed by level-set based deformable model LV 2D A4C Utilizing temporal coherence Nascimento (2010, 2013) DBN and particle filtering for dynamic modeling LV 2D A2C, A4C Jafari et al (2018) U-net and LSTM with additional optical flow input LV 2D A4C Utilizing unlabeled data Nascimento (2011, 2012) DBN on-line retrain using external classifier as additional supervision LV 2D A2C, A4C Smistad et al (2017) U-Net trained using labels generated by a Kalman filter based method LV and LA 2D A2C, A4C Yu et al (2017b) Dynamic CNN fine-tuning with mitral valve tracking to separate LV from LA Fetal LV 2D Jafari et al (2019) U-net with TL-net (Girdhar et al, 2016) based shape constraint on unannotated frames LV 2D A4C Utilizing data from multiple domains Chen et al (2016) FCN trained using annotated data of multiple anatomical structures Fetal head and LV 2D head, A2-5C Trained directly on large datasets Smistad et al (2018) Real filtering) framework with a transition model, in which the segmentation of the current cardiac phase depends on previous phases. The results show that this approach performs better than the previous method which does not take temporal information into account.…”

“…different 2D ultrasound views with various anatomical structures) can also help to improve the segmentation in one particular domain. Chen et al (2016) proposed a novel FCNbased network to utilize multi-domain data to learn generic feature representations. Combined with an iterative refinement scheme, the method has shown superior performance in detection and segmentation over traditional database-guided method (Georgescu et al, 2005), FCN trained on single-domain and other multi-domain training strategies.…”

Recurrent Aggregation Learning for Multi-view Echocardiographic Sequences Segmentation

“…The application scenario of existing methods is always limited and only suitable under a specific situation. They mostly focus on specific view [4] or single frames (i.e., without considering the sequence) [5] or one single vendor and center [6]. As for sequence segmentation, existing methods try to leverage temporal information by using a deformable model combined with the optical flow [7,8] or fine-tuning pretrained CNN dynamically with first frame's label till the last frame [9].…”