Yajing Zhang1 and Qitong Hu2
1Philips Healthcare, Suzhou, China, 2Beijing Longleding Medical Technology Co., Ltd., Beijing, China
Synopsis
MR has been widely used for
the diagnosis of hepatic hemangioma and cyst due to its significance of
detection on small lesions. This study proposes a deep learning based method to
detect the lesion volume in a three-dimensional manner on the dynamic
contrast-enhanced MR images with hemangioma and/or cyst lesions. The results
show good alignment of automated detection contour with the manually labelled lesion
contour by professional radiologists, as well as accurate classification of
lesion types.
Introduction
MR has been widely used for
the diagnosis of hepatic hemangioma and cyst due to its significance of
detection on small lesions [1]. However, the detection and delineation of the
lesions were labor-intensive with manual labelling and it heavily relies on the
clinical experience of the radiologist. The use of machine learning based
method for lesion detection has been increasely reported, with a few studies
using a 2D deep learning neural network [2], however, the 2D lesion ROIs may
mislead some regions interpretation due to the lack of cross-slice information.
In this study, we proposed a deep learning based method to detect the lesion
volume in a three-dimensional manner on the dynamic contrast-enhanced MR images
with hemangioma and/or cyst lesions. Method
A cohort of 61 patients
with either hepatic hemangioma or hepatic cyst were included in this study.
Clinical routine dynamic contrast-enhanced (DCE) MR imaging were acquired on a
3T scanner (Philips Healthcare, the Netherlands). The delayed phase MR images
were used to establish our deep learning neural network. Each patient data
contains 40-45 slices depending on the coverage of the liver. Manual labeling
of the lesions were performed by experienced radiologists with cross-check.
To leverage the correlative
information of lesion between consecutive slices, we employed a UNet neural
network architecture with 3D Convolution kernels [3]. The UNet was a symmetric
architecture in which the corresponding layer pairs in convolution and
deconvolution steps had the same size of feature map and there was a skip layer
structure between each pair as shown in Fig.1. The skip layer here passed the
low level subtle details from original image (feature map) which promoted the
segmentation result, while 3D kernels was preferred to 2D ones which could help
to capture the inter-layer correlations of anatomical information that provided
the lesion segmentation result a reasonable gain. The network was implemented
through Tensorflow [4], with GPU 1.4GHz, 1080Ti.
Original data were pre-processed by the DLTK software package. 80% of patient data were randomly arranged and used as training set and the rest as
validation set. Results
Figure 2 showed the comparison of lesion detection by
manual delineation vs. 3D UNet segmentation. Region in red showed the manual
labeled lesion, region in blue as the automated segmentation by our proposed
method. Consecutive slices with the lesion contoured by manual delineation vs.
by 3D UNet segmentation were shown. The region in purple indicated the overlap
between the two labels. Dice overlap ratio for the whole 3D lesion volume was
91%. Similarly, Figure 3 displayed the comparison of 3D lesion volume detection versus the manual delineation, and the dice overlap ratio of 84% were obtained. For the validation set, the average dice overlap ratio could be 78%, indicating a good alignment of the detecting lesion contours.Discussion and Conclusion
In this study, we employed a 3D convolutional neural network with skip layer which takes inter-layer information and low-level details into consideration. The overlap ratio on a volume lesion between our result and the manual labelling were 84% for abovementioned dataset. Compared to the 2D convolutional network such as in [2], this network takes the inter-slice information into account, which potentially reduce the misleading of suspicious regions which may appear on isolated slice. As more emerging methods of semantic segmentation in the field of computer vision have been introduced to medical image segmentation, the current model could be further improved.
Acknowledgements
No acknowledgement found.References
[1] Sasaki K, Ito K, Koike S, et al. Differentiation between
hepatic cyst and hemangioma: additive value of breath-hold, multisection
fluid-attenuated inversion-recovery magnetic resonance imaging using
half-Fourier acquisition single-shot turbo-spin-echo sequence. J Magn Reson
Imaging. 2005;21: 29-36.
[2] Zhang Y, et al. An automated lesion detection method on
hepatic hemangioma and hepatic cyst using fully convoluted network. ISMRM 2018.
[3] Erden B, Gamboa N, Wood S. 3D Convolutional Neural Network for Brain Tumor Segmentation. http://cs231n.stanford.edu/reports/2017/pdfs/526.pdf.
[4] Wongsuphasawat K, Smilkov D, Wexler J, et al. Visualizing Dataflow Graphs of Deep Learning Models in TensorFlow. IEEE Trans Vis Comput Graph. 2017;