Vishwa Sanjay Parekh1, Katarzyna J Macura2,3, and Michael A Jacobs2,3
1Department of Computer Science, Johns Hopkins University, Baltimore, MD, United States, 2The Russell H. Morgan Department of Radiology and Radiological Science, Johns Hopkins University School of Medicine, Baltimore, MD, United States, 3Sidney Kimmel Cancer Center, Johns Hopkins University School of Medicine, Baltimore, MD, United States
Synopsis
Artificial intelligence(AI)
and deep learning techniques are increasingly being used in radiological
applications. The true potential of deep learning in MRI applications can only be
achieved by developing an AI that can learn the underlying MRI physics rather
than a task that is specific to an organ or a particular tissue pathology. To
that end, we developed and tested a multiparametric deep learning model capable
of tissue segmentation and characterization in both breast cancer and stroke.
Introduction
Deep learning and
other artificial intelligence(AI) techniques have begun to play an integral
part in many different aspects of modern society1. Computer assisted clinical and radiological decision
support have major impact from these technologies, with some excellent initial
results2-10. However, the applications to date have been narrow, in
that, they are restricted to a single task e.g. brain segmentation from
multiparametric MRI(mpMRI). The true potential of AI will be achieved when the
model is generalizable to multiple different tasks. One recent successful
development of general AI came from Google DeepMind where they achieved human
level performance in playing a large number of Atari games11. This study investigates the feasibility of a generalized
deep network for segmentation and classification of mpMRI in different organs
with contrasting tissue pathologies based on the underlying physical modeling
of the MRI physics. Methods
We developed and validated a multiparametric deep
learning(MPDL) breast MRI network for segmentation of breast lesions in 139
patients. Input into MPDL network were mpMRI of T1 and T2-weighted imaging,
diffusion weighted imaging(DWI), apparent diffusion coefficient (ADC) mapping and dynamic contrast Enhanced(DCE) derived
tissue signatures. Our segmentation MPDL network was constructed from stacked
sparse autoencoders(SSAE) with five hidden layers12-14.
The MPDL was trained and validated on the breast cancer dataset using k-fold
cross validation with Dice Similarity(DS) metric. To apply the breast MPDL
signatures to another mpMRI data set, we used stroke mpMRI of perfusion-,
diffusion-, time-to-peak, T1-,T2-weighted imaging (PWI, DWI, TTP, T1WI and T2WI, respectively) obtained
from five stroke patients at one, two or three time points(12 studies) after
stroke: acute(≤12 hours), subacute(24-168 hours), and chronic(>168 hours). The
phase resolution of PWI in test dataset was matched to the phase resolution of
the training dataset using a combination of phase sliding window and wavelet
decomposition15. The experiment was
set-up as follows: Using the breast signature model, we applied the trained
MPDL on the stroke dataset. Areas of
infract and tissue at risk were segmented and tested against the MPDL results. The
DS between the MPDL and Eigenfilter(EI) segmented stroke lesions was employed
as the evaluation metric for validation of the MPDL segmentation stroke.
Statistical significance was set at p≤0.05Results
The MPDL
approach accurately defined
and segmented both the breast and stroke data. Fig. 1 demonstrates the
segmentation MPDL results on two malignant and two benign patients. The DS
index between the lesion defined by EI and MPDL segmentation demonstrated
excellent overlap with 0.87±0.05 for malignant patients and 0.85±0.07 for
benign patients. More importantly, the results from MPDL tested on the
independent stroke data were excellent as shown in fig. 2. There was high
correlation between the quantitative ADC and TTP values corresponding to the
lesion areas segmented by EI and MPDL (Rarea = 0.99, RTTP
= 0.99, RADC=0.86) corresponding to infract and tissue at risk. Fig. 3 demonstrates the scatterplot and the Bland
Altman plot between the quantitative values of ADC and TTP, and the lesion
areas segmented by the EI and MPDL algorithms. Furthermore, the percentage
difference in the lesion areas, TTP and ADC values segmented from EI and MPDL
were 6±2%,3±3%, and 7±7% respectively. Discussion
The knowledge transfer between the MPDL trained on
breast cancer and applied to stroke was excellent. These results demonstrate
the robustness of the MPDL tissue signature model based on the underlying MR characteristics
of each MRI parameter. For example, the T1 and T2 of the breast and stroke data
were similar. The DWI and ADC mapping were the same, except for the number of b
values. However, the MPDL method was able to adapt and apply the correct
signatures. Deep learning has previously been successfully applied to segment
and classify tissue types from brain, breast, prostate and other organs2-10. We successfully
demonstrated the AI capability of deep learning algorithms to learn the
underlying MRI physics. Our MPDL tissue signature model was able to segment different
tissue types from pathologies (stroke) that the MPDL had never “seen” before. Conclusion
The MPDL model was able to accurately segment the
tissue irrespective of the underlying organ or tissue pathology. Furthermore, the
MPDL model reveals the possibility of universal deep learning model that can
identify patterns in any organ or pathology or can be further personalized to
different applications.Acknowledgements
Funding: R01CA190299, 5P30CA006973 (IRAT), U01CA140204, U01CA166104, U01CA151235 and equipment grant from Nvidia Corporation.References
- LeCun Y, Bengio Y, Hinton G. Deep
learning. Nature. 2015;521(7553):436-444.
- Avendi
MR, Kheradvar A, Jafarkhani H. A combined deep-learning and deformable-model
approach to fully automatic segmentation of the left ventricle in cardiac MRI. Med Image Anal. May 2016;30:108-119.
- González AM.
Segmentation of Brain MRI Structures with Deep Machine Learning. MS
thesis. Universitat Politècnica de Catalunya, 2012.
- Guo Y, Gao Y,
Shen D. Deformable MR Prostate Segmentation via Deep Feature Learning and
Sparse Patch Matching. IEEE Trans Med
Imaging. Apr 2016;35(4):1077-1089.
- Li R, Zhang W,
Suk H-I, et al. Deep learning based imaging data completion for improved brain
disease diagnosis. International Conference on Medical
Image Computing and Computer-Assisted Intervention, 2014.
- Liu
S, Liu S, Cai W, Pujol S, Kikinis R, Feng D. Early diagnosis of Alzheimer's
Disease with deep learning. 2014 IEEE 11th International
Symposium on Biomedical Imaging (ISBI)2014.
- Shin HC, Orton
MR, Collins DJ, Doran SJ, Leach MO. Stacked Autoencoders for Unsupervised
Feature Learning and Multiple Organ Detection in a Pilot Study Using 4D Patient
Data. Ieee Transactions on Pattern
Analysis and Machine Intelligence. Aug 2013;35(8):1930-1943.
- Suk
HI, Lee SW, Shen D, Alzheimer's Disease Neuroimaging I. Latent feature
representation with stacked auto-encoder for AD/MCI diagnosis. Brain structure & function. Mar
2015;220(2):841-859.
- Xu J, Xiang L,
Liu Q, et al. Stacked Sparse Autoencoder (SSAE) for Nuclei Detection on Breast
Cancer Histopathology Images. IEEE Trans
Med Imaging. Jan 2016;35(1):119-130.
- Zhang W, Li R,
Deng H, et al. Deep convolutional neural networks for multi-modality isointense
infant brain image segmentation. NeuroImage.
Mar 2015;108:214-224.
- Mnih V,
Kavukcuoglu K, Silver D, et al. Playing atari with deep reinforcement learning.
arXiv preprint arXiv:1312.5602. 2013.
- Bengio Y, Lamblin
P, Popovici D, Larochelle H. Greedy layer-wise training of deep networks. Advances in neural information processing
systems. 2007;19:153.
- Hinton GE,
Osindero S, Teh YW. A fast learning algorithm for deep belief nets. Neural Comput. Jul 2006;18(7):1527-1554.
- Hinton
GE, Salakhutdinov RR. Reducing the dimensionality of data with neural networks.
Science. Jul 28
2006;313(5786):504-507.
- Akhbardeh
A, Jacobs MA, Methods and systems for registration of radiological images. US
patent US90084622015.