Lun M. Wong1, Ann D. King1, and Qiyong Ai1
1Department of Imaging and Interventional Radiology, The Chinese University of Hong Kong, Shatin, Hong Kong
Synopsis
Benign hyperplasia is a common finding in
the adenoid and walls of the nasopharynx and may hamper the detection of
early-stage nasopharyngeal carcinoma (NPC) on MRI. In this study we aim to utilize
deep learning to discriminate early-stage NPC from benign hyperplasia using T2-weighted-fat-suppressed
MR images. We tested our method on a dataset of 413 cases, comprising 203 with
early-stage NPC confined to the nasopharynx and 210 with benign hyperplasia. After
training with validation (n=350 and n=13 respectively) followed by testing (n=50),
the network achieved a promising result with a sensitivity of 100% and specificity
of 83% for NPC detection.
Introduction
Nasopharyngeal carcinoma (NPC) is a disease
prevalent in areas of the world such as South Asia China. Diagnosis of the
disease at an early-stage is crucial to the survival of patients with NPC. In
this regard, MRI is being used increasingly to detect small cancers that cannot
be seen through the endoscope1, especially in those patients who have been screened for the
disease using blood markers for the Epstein Barr Virus2,3.
However, benign hyperplasia of the
nasopharynx is common and causes enlargement of the adenoid and nasopharyngeal wall
thickening which may overlap in appearance with that of early-stage NPC on the
anatomical MRI sequences4 as shown in Fig. 1. Therefore, we have been researching new
techniques that may aid discrimination, including the diffusion weighted
imaging (DWI)5, but results showed some limitations. Furthermore, manual
delineations are required which is laborious and time consuming. Therefore, an automatic solution which can
discriminate benign and malignant nasopharyngeal lesions would be advantageous
for NPC detection.
Artificial neural
networks have the potential to perform this task.
While artificial neural networks have unparalleled natural image classification
capability the usage in medical imaging, especially for 3-dimensional (3D)
volumes, is still being explored. The purpose of this study is to investigate
whether convolutional neural networks (CNN) can be exploited to automatically
detect early-stage NPC on MR imaging.Method
This retrospective study was performed with
local institutional board approval. Patients were ethnically Chinese who underwent
T2-weighted-fat-suppressed MRI of the nasopharynx imaging in our
institution between 2010 and 2019. Two groups of patients were included in the
study: (1) those with newly biopsy-proven undifferentiated NPC whose primary
tumours were staged as T1 with a minimal axial diameter ≥ 5mm and (2) those
with benign
hyperplasia with a minimal axial diameter ≥ 5mm without
evidence of NPC on MRI and endoscopic examination. A total of 413 patients, 203
with NPC stage T1 and 210 with benign hyperplasia, were analysed.
The image volume went through a pipeline of
pre-processing filters, including alignment with the axis of symmetry using
EROS6, intensity Nyul normalization7 and cropping to uniform size of 20x444x444px.
We adopted the Residual Attention Network
(RAN)8 as our backbone and modified the convolutional kernels to suit the
3D inputs. Specifically, all the pooling layers were set to pool along 2D sagittal-coronal
layers only, whereas all convolutional layers convolved with 3D kernels.
Dropout layers were added to residual units to facilitate model robustness. The
RAN consists of an attention branch that allows the network to learn a feature
mask that reinforces useful deep features to mitigate the complex anatomy at
the nasopharynx.
We randomly divided our dataset into training
(n=350), validation (n=13) and testing (n=50) group. The network was allowed to
train for 5500 epochs over the training data with an initial learning rate of
10-4, which decayed exponentially after each epoch, optimizing
cross-entropy loss with stochastic gradient decent. The validation accuracy suggests
the network converges at approximately 2000 epochs as shown in Fig. 2. The
learnt model was then applied to the testing set. The sensitivity, specificity,
positive predictive value (PPV) and negative predictive value (NPV) were
evaluated.Results
There were five false positive and no false
negative results for NPC. The performance of the network showed a sensitivity of
100%, specificity of 83%, PPV of 81%, NPV of 100% and accuracy of 90%. The
computed probabilities of NPC and benign hyperplasia in the 50 patients in the
testing group are listed in Table 1.Discussion
Our results suggest CNN with the RAN architecture
as described is a promising tool for NPC detection on T2-weighted-fat-suppressed
MR and most cases were classified correctly with a high level of confidence,
indicated by the large prediction probability differences. Furthermore, three
of the five benign hyperplasia cases misclassified as NPC showed results with relatively
low confidence intervals. This suggests the probability ratio could alert the
radiologist to cases likely to be false positive for NPC and in the future
these cases could be reduced by tuning the threshold probability ratio.Conclusion
CNN is a promising tool for MRI detection
of early-stage nasopharyngeal carcinoma.Acknowledgements
References
-
King AD, Vlantis
AC, Bhatia KSS, et al. Primary nasopharyngeal carcinoma: diagnostic accuracy of
MR imaging versus that of endoscopy and endoscopic biopsy. Radiology.
2011;258(2):531-537.
-
Chan KCA, Woo
JKS, King A, et al. Analysis of Plasma Epstein-Barr Virus DNA to Screen for
Nasopharyngeal Cancer. N Engl J Med. 2017;377(6):513-522.
-
King AD, Woo JKS,
Ai QY, et al. Complementary roles of MRI and endoscopic examination in the
early detection of nasopharyngeal carcinoma. Ann Oncol.
2019;30(6):977-982.
-
King AD, Wong
LYS, Law BKH, et al. MR Imaging Criteria for the Detection of Nasopharyngeal
Carcinoma: Discrimination of Early-Stage Primary Tumors from Benign
Hyperplasia. AJNR Am J Neuroradiol. 2018;39(3):515-523.
-
Ai Q-Y, King AD,
Chan JSM, et al. Distinguishing early-stage nasopharyngeal carcinoma from
benign hyperplasia using intravoxel incoherent motion diffusion-weighted MRI. Eur
Radiol. 2019;29(10):5627-5634.
-
Chen X, Tian J,
Zhang Y, Yang X. A Robust Orientation Estimation Algorithm for Low Quality
Fingerprints. In: IWBRS 2005: Advances in Biometric Person Authentication.
Heidelberg: Springer; 2005:95-102.
-
Sun X, Shi L, Luo
Y, et al. Histogram-based normalization technique on human brain magnetic
resonance images from different acquisitions. Biomed Eng Online.
2015;14:73.
-
Wang F, Jiang M,
Qian C, et al. Residual attention network for image classification. In: Proceedings
- 30th IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2017.
Vol 2017-Janua. ; 2017:6450-6458.