Eleni Chiou1,2, Francesco Giganti3,4, Shonit Punwani5, Iasonas Kokkinos2, and Eleftheria Panagiotaki1,2
1Centre of Medical Imaging Computing, University College London, London, United Kingdom, 2Department of Computer Science, University College London, London, United Kingdom, 3Department of Radiology, UCLH NHS Foundation Trust, University College London, London, United Kingdom, 4Division of Surgery & Interventional Science, University College London, London, United Kingdom, 5Centre for Medical Imaging, Division of Medicine, University College London, London, United Kingdom
Synopsis
The successful adoption of convolutional neural networks (CNNs) for improved
diagnosis can be hindered for pathologies and clinical settings where the
amount of labelled training data is limited. In such cases, domain adaptation
provides a viable alternative. In this work we propose domain
adaptation to enhance the performance of prostate lesion segmentation on
VERDICT-MRI utilising diffusion weighted (DW)-MRI data from multi-parametric
(mp)-MRI acquisitions. Experimental results show that domain adaptation
significantly improves the segmentation performance on VERDICT-MRI.
Introduction
Convolutional neural
networks bring advances in many recognition tasks in medical
imaging. However, their performance relies on two conditions: i) the
availability of a significant amount of labelled training data and ii) whether
the training and test data belong in the same image domain. Satisfying both
conditions is challenging in biomedical applications due to the cost of human
labour and expertise required to produce ground truth labels, and the variations in image acquisition. In this work we tackle this
problem for prostate cancer diagnosis and an advanced diffusion weighted (DW)-magnetic resonance imaging (MRI) method called VERDICT. VERDICT-MRI is a non-invasive imaging technique
for cancer microstructure characterisation [1, 2, 3]. The method has been
recently in clinical trial to supplement standard multi-parametric (mp)-MRI for
prostate cancer diagnosis [4]. Compared to the naive DW-MRI from mp-MRI,
VERDICT-MRI has a richer protocol to probe the underlying microstructure and
reveal changes in tissue features similar to histology. However, the limited
amount of available labelled training data does not allow the training of robust
deep neural networks that could directly exploit the information in the raw VERDICT-MRI.
In cases where the amount of labelled training data is limited, domain
adaptation (DA) provides a viable solution. In this work, we investigate the use of DA for
lesion segmentation on raw VERDICT-MRI
data. Specifically, we use residual adapters (RAs) [5] and DW-MRI from mp-MRI
to train a robust network for lesion segmentation on VERDICT-MRI. Methods
Segmentation
architecture
The main component of our framework is an encoder-decoder network
for lesion segmentation [6, 7]. Given
an input image $$$\mathbf{X}\in\mathbb R^{H\times W\times C}$$$, where $$$C$$$ the number of input channels, the network provides a
segmentation mask
$$$\widehat{\mathbf{Y}}\in\mathcal Y^{H\times W}$$$, where
$$$\mathcal{Y}=\{0, 1\}$$$, indicating the class of each pixel in the image. We train the
network using as objective function a soft generalisation of the dice score
proposed in [8] and expressed as
$${\mathcal L_{DSC}} = \frac{2\sum_{k=1}^{K}{p_k g_k}}{\sum_{k=1}^{K}{p_k^2 +\sum_{k=1}^{K}{g_k^2 }}}, $$ where $$$K = H\times W$$$ is the number of voxels in the input images, $$$p_k\in[0,1]$$$ is the probability of voxel $$$k$$$ belonging in class $$$1$$$ and $$$g_k\in\{0,1\}$$$ is the ground truth label of voxel $$$k$$$.
Residual adapters
Let $$$\phi_l$$$ be a convolutional layer in the source domain network and $$$\mathbf F_l\in\mathbb R^{k\times k\times C_i\times C_o}$$$ be a set of filters for that layer, where $$$k\times k$$$ is the kernel size and $$$C_i$$$, $$$C_o$$$ are the number of input and output feature channels respectively. Let also $$$\mathbf G_l\in\mathbb R^{1\times1 \times C_i \times C_o}$$$ be a set of residual adapter filters installed in parallel with the existing set of filters $$$\mathbf F_l$$$. Given an input tensor $$$\mathbf x_l \in \mathbb R^{H \times W \times C_i}$$$, the output $$$\mathbf y_l \in \mathbb R^{H\times W\times C_o}$$$ of layer $$$\phi_l$$$ is given by $$\mathbf{y}_l= \mathbf{F}_l*\mathbf{x}+\mathbf{G}_l*\mathbf{x}.$$
Introducing RAs in the pre-trained network ensures that most parameters stay the same, but also that the new unit introduces a small, but effective modification that can effectively compensate for changes in the feature statistics.Materials
VERDICT-MRI
We use
VERDICT-MRI data from $$$60$$$ men. We acquire VERDICT-MRI images with
pulsed-gradient spin-echo sequence (PGSE) using an optimised imaging protocol
for VERDICT prostate characterisation with $$$5$$$ b-values ($$$90-3000\ \rm{s/mm^2}$$$), in $$$3$$$ orthogonal directions, on a $$$3$$$T scanner [10].
We also acquire images with $$$b=0\ \rm{s/mm^2}$$$ before each b-value acquisition. VERDICT-MRI data was registered
using rigid registration [11]. A dedicated radiologist contoured the lesions (Likert
score of $$$3$$$ and higher) on VERDICT-MRI using mp-MRI for guidance.
DW-MRI from
mp-MRI acquisitions
We use DW-MRI data from the ProstateX challenge dataset consisting
of training mp-MRI data acquired from $$$204$$$ patients [12]. The DW-MRI data were
acquired with a single-shot echo planar imaging sequence with diffusion
encoding gradients in three directions. Three b-values were acquired and
subsequently, the ADC map and a b-value image at $$$b=1400\ \rm{s/mm^2}$$$ were
calculated by the scanner software. We use DW-MRI data from $$$114$$$ patients. Since
the ProstateX dataset provides only the position of the lesion, a radiologist
manually annotated the lesions on the ADC map using as reference the provided
position of the lesion. Results
We evaluate
the performance of RAs against two baselines. Below we
describe the experiments we perform.
- Residual adapters (RA): We introduce the RAs in the pre-trained
network and update only the RAs using VERDICT-MRI data.
- Pre-trained network (PT): We perform standard fine-tuning of the pre-trained network using VERDICT-MRI data.
- VERDICT-MRI data (VERDICT):
We train the encoder-decoder network from scratch in the target domain.
We employ a
nested $$$5$$$-fold cross validation to select the hyper-parameters (dropout value,
decay rate) and to evaluate the performance. We report the average performance (average recall, precision, dice similarity coefficient (DSC), and average precision (AP)) across the outer $$$5$$$-fold cross validation. We achieve the
best performance when we introduce RAs in the
pre-trained network (Table 1). Figure 2 shows on a $$$b=2000\ \rm{s/mm^2}$$$ image qualitative results
produced by the models.
Conclusion
In this work
we employed DA to leverage labelled DW-MRI data from mp-MRI acquisitions to
improve prostate lesion segmentation on VERDICT-MRI. Experimental results
indicate that introducing RAs in a pre-trained network in the source domain
significantly improves the performance in the target domain.
Acknowledgements
This research is
funded by EPSRC grand EP/N021967/1. The Titan Xp used for this research was
donated by the NVIDIA Corporation.References
1. Panagiotaki, E., et al.: Noninvasive quantification of solid tumor
microstructure using VERDICT-MRI. Cancer Research 74, 7, 2014.
2. Panagiotaki, E., et al.: Microstructural characterization of
normal and malignant human prostate tissue with vascular, extracellular, and
restricted diffusion for cytometry in tumours magnetic resonance imaging.
Investigative Radiology 50, 4, 2015.
3. Johnston, W., E., et al.: Verdict MRI for prostate cancer:
Intracellular volume fraction versus apparent diffusion coefficient. Radiology,
291, 2, 2019.
4. Johnston, W., E., et al.: INNOVATE: A prospective cohort study
combining serum and urinary biomarkers with novel diffusion-weighted magnetic
resonance imaging for the prediction and characterization of prostate cancer. BMC
Cancer, 16, 816, 2016.
5. Rebuffi, S.A., et al.: Efficient parametrization of multi-domain
deep neural networks. In: CVPR, 2018.
6. Badrinarayanan, V., et al.: Segnet: A deep convolutional
encoder-decoder architecture for image segmentation. TPAMI, 2017.
7. Chen, L.C., et al.: Encoder-decoder with atrous separable
convolution for semantic image segmentation. In: ECCV, 2018.
8. Milletari, F., et al.: V-net: Fully convolutional neural networks
for volumetric medical image segmentation. In: 3DV, 2016.
9. He, K., et al.: Deep residual learning for image recognition. In:
CVPR, 2016.
10. Panagiotaki, E., et al.: Optimised VERDICT MRI protocol for
prostate cancer characterisation. In: ISMRM, 2015.
11. Ourselin, S., et al.: Reconstructing a 3D structure from serial
histological sections. IVC 19, 1, 2001.
12. Litjens, G., et al.: Computer-aided detection of prostate cancer
in MRI. TMI 33, 5, 2014.
13. Paszke, A., et al.: Automatic differentiation in pytorch. In: Autodiff
Workshop, NIPS, 2017.
14. Srivastava, N., et al.: Dropout: A simple way to prevent neural
networks from overfitting. JMLR 15, 2014.