0% found this document useful (0 votes)
5 views8 pages

Elsiver 7

This study presents a computer-aided detection (CAD) algorithm utilizing Convolutional Neural Networks (CNN) to identify ureteral stones in high-resolution computed tomography (CT) scans. The proposed method directly analyzes raw CT data, achieving a sensitivity of 100% with an average of 2.68 false positives per patient from a dataset of 465 scans. The research addresses challenges in distinguishing stones from similar structures and incorporates anatomical information to enhance detection accuracy.

Uploaded by

keerthi.080403
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
0% found this document useful (0 votes)
5 views8 pages

Elsiver 7

This study presents a computer-aided detection (CAD) algorithm utilizing Convolutional Neural Networks (CNN) to identify ureteral stones in high-resolution computed tomography (CT) scans. The proposed method directly analyzes raw CT data, achieving a sensitivity of 100% with an average of 2.68 false positives per patient from a dataset of 465 scans. The research addresses challenges in distinguishing stones from similar structures and incorporates anatomical information to enhance detection accuracy.

Uploaded by

keerthi.080403
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 8

Computers in Biology and Medicine 97 (2018) 153–160

Contents lists available at ScienceDirect

Computers in Biology and Medicine


journal homepage: www.elsevier.com/locate/compbiomed

Computer aided detection of ureteral stones in thin slice computed


tomography volumes using Convolutional Neural Networks
Martin L€angkvist a, *, Johan Jendeberg b, Per Thunberg c, Amy Loutfi a, Mats Liden b
a €
Center for Applied Autonomous Sensor Systems, Orebro €
University, Orebro, Sweden
b €
Department of Radiology, Faculty of Medicine and Health, Orebro €
University, Orebro, Sweden
c € €
Department of Medical Physics, Faculty of Medicine and Health, Orebro University, Orebro, Sweden

A R T I C L E I N F O A B S T R A C T

Keywords: Computed tomography (CT) is the method of choice for diagnosing ureteral stones - kidney stones that obstruct
Computer aided detection the ureter. The purpose of this study is to develop a computer aided detection (CAD) algorithm for identifying a
Ureteral stone ureteral stone in thin slice CT volumes. The challenge in CAD for urinary stones lies in the similarity in shape and
Convolutional neural networks intensity of stones with non-stone structures and how to efficiently deal with large high-resolution CT volumes.
Computed tomography We address these challenges by using a Convolutional Neural Network (CNN) that works directly on the high
Training set selection
resolution CT volumes. The method is evaluated on a large data base of 465 clinically acquired high-resolution CT
False positive reduction
volumes of the urinary tract with labeling of ureteral stones performed by a radiologist. The best model using 2:5D
input data and anatomical information achieved a sensitivity of 100% and an average of 2.68 false-positives per
patient on a test set of 88 scans.

1. Introduction While previous works mainly have analyzed 2D images by extracting


features from segmented regions, this work instead operates on the 3D
During the last decades, computed tomography (CT) has emerged as volumes directly on the raw pixel data using a Convolutional Neural
the method of choice for diagnosing ureteral stones [1]. The formation of Network (CNN) [8,9]. Pixel/voxel-based machine learning (PML) has
crystals in the urinary system - urinary stone disease - is a common recently emerged in medical image processing as an alternative to
condition with a lifetime risk of approximately 10% [2]. A ureteral stone traditional feature-based ML [10]. The advantage of PML methods is that
is a kidney stone that has been displaced and passed into and thereby the algorithms learn directly from the raw data instead of using extracted
obstructing the ureter causing severe pain. An obstructing ureteral stone features from segmented objects and thus no information is lost and, in
can further be complicated by infection and renal failure. addition, removes the need for researching, extracting, and selecting
Small ureteral stones are likely to pass spontaneously while larger features and segmentation methods.
stones may need interventional therapy such as extracorporeal shock A challenge with PML methods for high-resolution CT data is how to
wave lithotripsy or endoscopic lithotripsy. Diagnosis, localization and efficiently train the model and how to deal with the large amount of data.
estimation of prognosis for spontaneous passage of ureteral stones are In this work, we propose a method that operates directly on the raw data
performed by a radiologist by reviewing the cross-sectional slices of the of selected volumes-of-interest (VOI) and present a strategy for reducing
CT scan, that produces a grayscale 3D volume of the examined body [3]. the number of false positives using anatomical information, which is a
Computer assisted detection (CAD) tools for assisting radiologists are challenge for many CAD systems for medical applications where a false
gradually being introduced in the everyday image review [4]. A few negative is more critical than a false positive.
previous works related to CAD in urinary stone disease have focused on This paper has the following contributions:
detection of kidney stones that have not passed into the ureters [5,6], and
image features for discrimination among already detected calcifications  Performs ureteral stone detection from raw high-resolution CT scans
between ureteral stones and mimickers [7]. To the best of our knowledge using Convolutional Neural Networks
there are no previous reports of CAD systems for the more complex task  Proposes a probabilistic approach to include anatomical information
of detecting ureteral stones. in the decision process

* Corresponding author.
E-mail address: [email protected] (M. L€angkvist).

https://doi.org/10.1016/j.compbiomed.2018.04.021
Received 2 February 2018; Received in revised form 6 April 2018; Accepted 23 April 2018
0010-4825/© 2018 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-
nc-nd/4.0/).
M. L€
angkvist et al. Computers in Biology and Medicine 97 (2018) 153–160

 Present a training set selection strategy using similarity ranking therefore require fewer parameters to learn and takes advantage of the
 Evaluates various design choices and model parameters for the pro- spatial structure in the input data. The CNN is also capable of learning
posed method slightly translational and rotational invariant features (also called filters),
which is a benefit for structured data.
The rest of the paper is organized as follows. The CT scans and the An overview of how a CNN is used to classify a CT volume is seen in
proposed method is described in Section 2. The experimental results is Fig. 1. The steps of a single-layer CNN consists of a convolving layer,
performed in Section 3. Finally, comparison with related work and followed by a non-linear activation function, and finally subsampling
conclusions are given in Section 4. with pooling. A fully-connected layer (FC-layer) is connected to the
output from the pooling layer, followed by a classifier attached to the FC-
2. Materials and method layer for the final classification. The model parameters to be trained are
θ ¼ fW1 ; …; Wk ; b1 ; …; bk ; Ws ; bs g and the hyperparameters to set are the
2.1. Input data input size m, filter size n, pooling dimension p, and number of filters k.
The selection of these parameters is further discussed in Section 3.3.
The full data set consists of 465 clinically acquired unenhanced The convolutional layer fijk for filter k for an input image I c with c
abdominal CT scans with 0:8  0:8  1 mm voxel size. The study group number of channels is computed as:
included both men and women  18 years old with one scan per patient. !
The CT scans were obtained with three different CT systems (Brilliance, XX
n1 X
n1

Philips Medical Systems Best, The Netherlands; Somatom Definition fijk ¼ σ bk þ Wkabc Iiþa;jþb
c
(1)
c a¼0 b¼0
Flash, Siemens, Erlangen, Germany; Somatom Definition AS, Siemens,
Erlangen, Germany). All scans were obtained with an intermediate dose where Wi and bi are the filter and bias for the ith filter, n is the filter size,
CT protocol for the urinary tract (120 kVp, CTDI 5mGy, no dose modu- _ is the activation function. The sigmoid activation function, σ ðxÞ ¼
and σ ðÞ
lation). The weight of the included patients is unknown and the noise
1=ð1 þ ex Þ, is used in this work.
depends on the body size since no dose modulation was used. A solitary
The pooling layer is calculated by taking the max over non-
ureteral stone is present in all of the 465 CT scans and the position has
overlapping subregions of the convolutional layer with pooling size p.
been provided by a radiologist. Each CT scan, I 2 ℕ3 , is 512  512 
The pooling layer gijk for the kth filter is calculated as:
400 pixels where the intensity c in pixel position ði; j; kÞ is defined by
I ijk ¼ c where the pixel intensities in the data sets are in the range c ¼ ½   
gkij ¼ max f1þpði1Þ:pi;1þpðj1Þ:pj
k
(2)
1024 3071. The intensity of each pixel represents the attenuation value
in Hounsfield Units (HU) scale and is related to the material composition The concatenated pooling layer is then used as input to a FC-layer.
at that location inside the body with air having an attenuation value of The dimension of the pooling layer depends on the input size m, filter
1000 HU and water 0 HU. size n, number of filters k, and pooling dimension p according to
The Regional Research Ethics Board approved the study protocol and  3
waived the informed consent requirement. k mnþ1
p
. The FC-layer has h number of hidden units, which we set to
h ¼ 100 in this work.
2.2. Convolutional Neural Networks Finally, a classifier is attached to the FC-layer for the final classifi-
cation. We use a softmax classifier which outputs a vector with K ele-
Convolutional Neural Networks (CNNs or ConvNets) [11] have been ments that represents the normalized classification probabilities for each
successful in computer vision tasks with structured multiple arrays such class K. In this work we set K ¼ 2 for the class of stone and non-stone. The
as object recognition in images [12], speech recognition from time-series probabilities are calculated as:
data [13], and videos [14].
 
A conventional Artificial Neural Network (ANN) consists of an input exp xT Wsj
Pðy ¼ jjxÞ ¼ PK   (3)
layer, hidden layer, and output layer. ANNs are fully connected, which
k¼1 exp x Ws
T k

means that each hidden unit has a connection to all input units and each
output unit has a connection to each hidden unit. For an input unit Ij , bias 2.2.1. Calculating probability map
unit bi , and a weight matrix Wij , the hidden unit hi is calculated as hi ¼ Locating a calcification within or outside the ureter is the key prin-
PJ
j¼1 Wij Ij þ bi . For high-dimensional inputs, such as images, where J is
ciple that radiologists use when differentiating between stones and
large, the number of parameters to learn in the weight matrix becomes phleboliths. When surrounding anatomical landmarks are removed, the
too high and the model suffers from the curse of dimensionality [15]. differentiation is more difficult.
A CNN, instead, uses local connections with tied weights and One approach to incorporating more information from the

Fig. 1. Process of classifying a volume-of-interest. The convolutional layer is formed by convolving the filters over the input volume and the pooling layer is a
subsampling of the convolutional layer. The input vector that is fed to the classifier is formed by concatenating all pooling layers for each filter.

154
M. L€
angkvist et al. Computers in Biology and Medicine 97 (2018) 153–160

Fig. 2. (a) Before scan alignment (b) After scan alignment. White regions show where the reference scan (magenta) and registered scan (green) have the same in-
tensities. Only translational and rotational transformation is used. The image shows the result projected in the frontal plane for clarity. Best seen in color. (For
interpretation of the references to color in this figure legend, the reader is referred to the Web version of this article.)

surrounding area is to use a multi-scale CNN that uses downsampled to the aligned stone positions. The initial mean for the GMM components
images [16]. Since we have a relatively large database of scans and an- are set using k-means. The probability map is then normalized in each
notated stone positions, we will instead calculate a probability map of the slice to get a probability distribution for each slice. Fig. 3 show the lo-
location of real stones and use this information in a post-processing step cations for all stones before and after alignment. The mean of the prob-
to filter out the number of false-positives. ability map in each dimension are shown on the walls and the floor. It can
Before calculating the probability map, the skeletal structures of all be seen that the alignment reduces the variance of the stone locations and
scans were first automatically aligned with a translational and rotational allows for a more accurate probability distribution of the locations of the
transformation in all three dimensions to a randomly selected reference stones.
scan. The ureters are located in the retroperitoneal space, stretching from
the kidneys cranially and down to the bladder. Skeletal registration was 3. Results
chosen, since the relation between the ureters in the retroperitoneal
space and the spine, is only to a minor degree dependent on body size. A The dataset consists of 465 clinically acquired unenhanced abdominal
non-deforming registration could therefore be used. CT scans of patients suffering from suspected renal colic. For training the
To achieve the skeletal registration, the scans and the reference scan CNN, the scans are randomly divided into 80% training set and 20% for
were pre-processed by thresholding the intensities to values above 100, testing after 28 of the scans that contained a too small or too large stone
then downsampled by a factor of 5, and then smoothed with a Gaussian were removed, resulting in 349 training scans and 88 testing scans. The
5  5  5 filter before calculating the alignment. Fig. 2 shows how one data pre-processing steps consist of selecting volumes-of-interest (VOI)
scan is aligned to the reference scan. The purpose of the alignment is to using connected components (CC) and a training set selection method
more accurately calculate a probability of the distribution of the locations that ranks the CCs.
of the stones.
The locations of all stones in the training set are tracked before and
3.1. Finding connected components
after the alignment. The probability map is calculated by fitting a
Gaussian Mixture Model (GMM) in three dimensions with 6 components
The process of finding connected components (CCs) in one scan

Fig. 3. Stone positions (a) before scan alignment, and (b) after scan alignment. The mean of the probability map in each dimension are shown on the walls and the
floor of the plot. The gray color represents the Z-value for easier visualization.

155
M. L€
angkvist et al. Computers in Biology and Medicine 97 (2018) 153–160

Fig. 4. Process of binarizing and finding connected components. Left image shows an example of an axial 2D section through the pelvis and the lower part of the
urinary tract. Right image shows the data after binarizing and the connected components.

consists of first binarizing the volume using the threshold HUthresh and have a minimum euclidean distance. Fig. 5(a) show one slice of a 25  25
then group all pixels using 6-connectivity in CCs. CCs with a number of patch of the raw data for all the stones in the training set and Figure 5(b)
voxels below volmin and above volmax were discarded from the analysis. show the same data for the non-stones in one scan sorted by the
The binarization threshold (HUthresh) was with no loss of general- similarity.
ization fixed to 250 HU since all common urinary stone types have In order to increase the number of training examples for the stones, a
considerably higher attenuation values [3]. The minimum stone volume maximum total of 10 randomized pixels in each CC for the stones are
(volmin) were chosen to 5 voxels and the maximum stone volume selected, as opposed to only one for the non-stone CCs. Each training
(volmax) were chosen to 500 voxels, corresponding approximately to the example for the stones are then mirrored in the YZ-plane. This results in
volume of a perfect sphere with 2 mm and 9 mm diameter, respectively. each scan giving 100 non-stone examples and at most 20 stone examples.
The limits were chosen since because virtually all stones smaller than
2 mm pass spontaneously without treatment, while stones with a short
axis larger than 9 mm are unlikely to be found in the ureter [17]. 3.3. Evaluating training schemes and hyperparameters
The effect of binarizing with the chosen threshold can be seen in
Fig. 4 where voxels belonging to soft tissue have been filtered out while There are a number of design choices and model hyperparameters
urinary stones and skeletal structures remain. that need to be set. The first is the choice of input data. In this work we
present three training schemes of using 2D, 2.5D, or 3D input data, see
Fig. 6. The input size, m, determines the dimensionality of the input data.
3.2. Training set selection The image size m used as input to the CNN is set among m ¼
½5; 11; 25. The filter size n and pooling dimension p are set among n ¼
The remaining number of connected components in each scan after ½3; 6; 11 and p ¼ ½1; 2; 5, respectively. The filter size and pooling
filtering out the ones that are too small or too large can vary between a dimension is set so that the size of the pooling layer is a 3  3  k. The
few hundreds to a few thousands. Each scan contains only one CC con- number of filters k is set among k ¼ ½20; 50 for each of the training
taining a ureteral stone and the rest are not ureteral stones. A subset of schemes.
the non-stones need to be selected to balance the training set. The The attenuation values in the CT scans range between 1024 HU and
strategy for selecting the training examples is to select all CCs with a 3071 HU. In order to decrease the training time of the CNN and avoid
stone and primarily select CCs that resemble a real stone for the non- getting stuck in early local optima, the values are normalized to values
stone examples. The process of selecting examples for non-stones is between 0 and 1 by first saturating the upper limit for the attenuation
done by calculating the euclidean distance between the raw data of a values to 1000 HU and then subtracting each pixel in the scan with a
cube of 25  25  25 around a random pixel in each non-stone CC to the minimum value and dividing by the difference between the maximum
same data for a random real stone and then selecting the top 100 CCs that and the minimum value according to IN ¼ ðI  minðIÞÞ=ðmaxðIÞ 

Fig. 5. (a) All stones in the training set. (b) All non-stones in one scan sorted by similarity, calculated with the euclidean distance, to another random real stone.

156
M. L€
angkvist et al. Computers in Biology and Medicine 97 (2018) 153–160

Fig. 6. Three different ways of inputting the data to the algorithm: (a) 2D, (b) 2.5D, and (c) 3D.

minðIÞÞ. 3.4. Classification accuracy


The training is done using supervised backpropagation with mini-
batch stochastic gradient descent (SGD), early-stopping and mo- Training was done by randomly divide the original 465 scans into 349
mentum. The batch size is selected as 100 and the learning rate is initially scans for training and 88 scans for testing after removing scans that
set to 0.1 and uses a decaying learning rate if the accuracy on the vali- contain a stone smaller than 5 or larger than 500. From each of the scan
dation set is not improved over a period of 10 epochs. in the training set, 100 examples of non-stone were extracted using the
Due to the class-imbalance in the training data, the training set is first method explained in Section 3.2 and 20 examples of stones. The best
balanced to have equal class distribution before each training epoch and input method and model architecture from Section 3.3 was selected,
the F1-score is used instead of classification accuracy for evaluation on namely the training scheme that uses 2:5D input data, image size m ¼ 11,
the full testing set. filter size n ¼ 6, pooling dimension p ¼ 2, and number of filters k ¼ 50.
The reported F1-score for the three different training schemes and The model parameters of the CNN was trained using supervised back-
varying number of hyperparameters on the testing set are shown in propagation and stochastic gradient descent with learning rate 0.1 until
Table 1. The highest accuracy with a small CNN model (20 filters) ach-
ieves the highest F1-score with the 2.5D scheme and with a small image
size of m ¼ 5. The highest accuracy with a larger CNN model (50 filters) Table 1
achieves the highest F1-score with the 2.5D scheme and with a medium F1-score for three training schemes and varying number of filters and image
image size of m ¼ 11. Both the 2:5D and 3D scheme were better than sizes. The highest F1-score for each image size is marked in bold.
using 2D input. A medium size input image of m ¼ 11 achieved a higher # filters (k) Image size (m)
F1-score than both the smaller and the larger image sizes for all schemes, m¼5 m ¼ 11 m ¼ 25
meaning that a larger input image size does not improve performance.
2D k ¼ 20 0.688 0.734 0.667
However, when a large image size of m ¼ 25 is used, the performance is k ¼ 50 0.626 0.720 0.714
increased when the number of filters is increased. This comes at a cost of 2.5D k ¼ 20 0.776 0.772 0.719
increased training time. A trade-off between classification performance k ¼ 50 0.755 0.783 0.735
and training time resulted in the choice of using the 2:5D scheme with 3D k ¼ 20 0.701 0.749 0.713
k ¼ 50 0.723 0.768 0.734
image size m ¼ 11 and number of filters k ¼ 20.

157
M. L€
angkvist et al. Computers in Biology and Medicine 97 (2018) 153–160

Table 2 For testing and evaluation, all pixels in each connected component
Confusion matrix for all CCs in the test set. F1-score 0.35, Recall 0.21, Precision that are between 5 and 500 pixels large are classified into stone or non-
1.00. stone and the final prediction is the average over all pixel predictions in
Predicted that CC. The result was that the model correctly predicted all 88 stones (1
Positive Negative
stone in each scan in the test set), 0 false-negatives, 325 false-positives,
and a total of 18829 true-negatives, see Table 2 for the confusion ma-
Actual Positive 88 0
trix. This gives an average of 3:7  4:9 false-positives per scan.
Negative 325 18829
A further analysis of the results can be seen in Fig. 7 were the number
of false-positives and true-negatives for each individual CT scan in the
convergence. test set are shown. The number of true-negatives are mostly constant with

Fig. 7. Number of false-positives (gray bars) and true-negatives (black line) of each scan in the test set.

Fig. 8. ROC curve for different sizes of CCs.

158
M. L€
angkvist et al. Computers in Biology and Medicine 97 (2018) 153–160

Fig. 9. FROC curve for different trade-offs between classification certainty and stone location probability. The least amount of false positives with a 100% sensitivity is
achieved with α ¼ 0:6 (black thick line).

an average of 214:0  264:0 per scan, except for scan number 128 that 3.69 false-positives per scan with sensitivity 100%, which is the same
has over 2000 true-negatives. Notice that this large amount of negatives result reported in Section3.4. The best trade-off is achieved with α ¼ 0:6
did not have a significant impact on the number of false-positives that which gives an AUC ¼ 0:9971 and 2.68 false-positives per scan with
was only 3 for this scan. The largest amount of false-positive was detected sensitivity 100%. The use of anatomical information reduced the amount
on scan number 101 with 25 false-positives. A fourth (25%) of the scans of false-positives by 27:4%.
have 0 false-positives and 40:9% of the scans have only 1 or 0 false-
positives. 4. Discussion and conclusion
Beside the individual scan, the size of the CC also have an influence of
the number of false-positives. Fig. 8 shows the ROC curve for different In the present study, a Convolutional Neural Network was used for
sizes of CCs in the test set. This difference is explained by the use of a detecting ureteral stones in thin slice CT scans. The main contribution of
small image size of 11  11 pixels and the surrounding information this work is to approach automatic detection of ureteral stones, by using a
around large stones are not captured. The experiments from Section 3.3 representational learning algorithm that is not dependent on the choice
showed that using a larger image size did not improve the overall ac- of features or segmentation methods, and that takes advantage of the
curacy since most stones are small. Therefore, we use the anatomical high-resolution CT data.
information of the location of the CC to reduce the number of false- Without using anatomical information the CNN model achieved a
positives for large-sized CCs. sensitivity of 100% and an average of 3.69 false-positives per patient on a
test set of 88 scans. The use of a probability distribution map of stone
3.4.1. Results on using probability map locations achieved a sensitivity of 100% and an average of 2.68 false-
The probability distribution map from Section 2.2.1 is used to change positives per patient.
the output prediction probabilities for the classification from the CNN Convolutional Neural Networks are rapidly becoming a popular
network. The probability map M gives a probability that the CC at method for medical image analysis [18,19]. They have previously been
location x is a stone by Pðy ¼ 1jMðxÞÞ and the CNN model gives a pre- used on a number of medical image modalities, including CT data for
diction probability that the CC with input data I is a stone by Pðy ¼ tasks such as brain tumor segmentation [16,20], pancreas segmentation
1jIðxÞÞ. The trade-off between how much of these prediction probabili- [21], and lesion segmentation in livers [22]. To the best of our knowl-
ties should influence the final decision is determined by the parameter α edge, there are no previously published works that uses CNNs on CT
by: scans for ureteral stone detection.
Most previous work for computer aided detection of renal stones use a
Pðy ¼ 1jIðxÞ; MðxÞÞ ¼ αPðy ¼ 1jIðxÞÞ þ ð1  αÞPðy ¼ 1jMðxÞÞ (4) work-flow of noise reduction, candidate selection, and then a feature-
based method for the final classification. The work by Liu et al. uses
where α is between 0 and 1.
total variation (TV) flow for image noise reduction and MSER features for
The results on a Free Response Operating Characteristic curve (FROC)
finding calculi candidates, and computes a total of 7 texture and shape
for different values of α can be seen in Fig. 9. The lowest value for the
features to train a Support Vector Machine (SVM) for the task of identi-
area-under-the-curve (AUC) is 0.605 and is obtained when α ¼ 0, i.e.,
fying renal calculi in CT scans [5]. The method was validated on a data
when only the prediction from the probability map is used and the pre-
set of 192 patients with a false positive rate of 8 per patient and a
diction from the CNN model is ignored. When α ¼ 1, i.e., only the pre-
sensitivity of 69%. The work by Ebrahimi et al. correctly predicted the
diction from the CNN is used, we get an AUC ¼ 0:9967 and on average of

159
M. L€
angkvist et al. Computers in Biology and Medicine 97 (2018) 153–160

presence of a kidney stone in 24 of 29 patients and the absence of a [2] O.W. Moe, Kidney stones: pathophysiology and medical management, The lancet
367 (9507) (2006) 333–344.
kidney stone in 9 of 10 patients using a method of localization and
[3] A.R. Kambadakone, B.H. Eisner, O.A. Catalano, D.V. Sahani, New and evolving
thresholding [6]. concepts in the imaging and management of urolithiasis: urologists' perspective,
A feature-based approach to differentiate ureteral stones and vascular RadioGraphics 30 (3) (2010) 603–623, https://doi.org/10.1148/rg.303095146
calcifications has previously been applied by training an Artificial Neural pMID: 20462984. arXiv, https://doi.org/10.1148/rg.303095146, https://doi.org/
10.1148/rg.303095146.
Network (ANN) on a total of 8 shape and texture features on non-contrast [4] K. Doi, Computer-aided diagnosis in medical imaging: historical review, current
CT images [7]. The method was validated on 59 ureter stones and 53 status and future potential, Computerized Med. Imag. Graph. 31 (4–5) (2007)
vascular calcifications and resulted in an area under the curve (AUC) of 198–211.
[5] J. Liu, S. Wang, E.B. Turkbey, M.G. Linguraru, J. Yao, R.M. Summers, Computer-
0.85 for the shape parameters and 0.88 for the texture parameters. The aided detection of renal calculi from noncontrast CT images using TV-flow and
more general problem of detecting the ureteral stones was not addressed MSER features, Medical Phys. 42 (1) (2015) 144–153.
in the study. [6] S. Ebrahimi, V.Y. Mariano, Image quality improvement in kidney stone detection on
computed tomography images, J. Image Graphics 3 (1) (2015) 40–46, https://
Our study differentiates from the previous works in several ways. doi.org/10.18178/joig.3.1.40-46.
First, the task in the present study is detection of ureteral stones, which is a [7] H.J. Lee, K.G. Kim, S.I. Hwang, S.H. Kim, S.-S. Byun, S.E. Lee, S.K. Hong, J.Y. Cho,
more challenging task for radiologists compared to stones in the kidney, C.G. Seong, Differentiation of urinary stone and vascular calcifications on non-
contrast ct images: an initial experience using computer aided diagnosis, J. Digit.
and a more general problem than differentiation between stones and Imag. 23 (3) (2010) 268–276.
other calcifications. Second, we use the raw pixels on 3D volumes instead [8] Y. LeCun, L. Bottou, Y. Bengio, P. Haffner, Gradient-based learning applied to
of extracted features and we also propose a method for using the document recognition, Proc. IEEE 86 (11) (1998) 2278–2324.
[9] Y. LeCun, B.E. Boser, J.S. Denker, D. Henderson, R.E. Howard, W.E. Hubbard,
anatomical information in the decision process.
L.D. Jackel, Handwritten digit recognition with a back-propagation network, in:
In comparison with previous works, the present study represents D.S. Touretzky (Ed.), Advances in Neural Information Processing Systems 2,
another step towards an automatic detection algorithm for ureteral Morgan-Kaufmann, 1990, pp. 396–404.
stones. However, from a clinical point-of-view, while keeping 100% [10] K. Suzuki, Pixel-based machine learning in medical imaging, Int. J. Biomed. Imag.
2012 (2012) 1 article ID 792079, https://doi.org/10.1155/2012/792079.
sensitivity, the false positive rate of 2.7 per CT scan is still not sufficient [11] Y. Lecun, L. Bottou, Y. Bengio, P. Haffner, Gradient-based learning applied to
for replacing the primary diagnostic reading by the radiologist. Radiol- document recognition, Proc. IEEE 86 (11) (1998) 2278–2324, https://doi.org/
ogists have a reported sensitivity and specificity of approximately 95  10.1109/5.726791.
[12] A. Krizhevsky, I. Sutskever, G.E. Hinton, Imagenet classification with deep
100% for ureteral stone CT [3,23]. The promising results in the present convolutional neural networks, Adv. Neural Inf. Process. Syst. 2 (2012) 1097–1105.
study clearly shows the possibilities of CNN for computer aided diagnosis [13] H. Lee, Y. Largman, P. Pham, A.Y. Ng, Unsupervised feature learning for audio
in CT images. With further improvement of the methods, the develop- classification using convolutional deep belief networks, in: Proceedings of the 22Nd
International Conference on Neural Information Processing Systems, Curran
ment of an automatic detection algorithm for ureteral stones ready for Associates Inc., USA, 2009, pp. 1096–1104. NIPS’09, http://dl.acm.org/citation.
clinical use, is within reach. Future work include methods for reducing cfm?id¼2984093.2984217.
the amount of pre-processing steps such as finding connected compo- [14] G.W. Taylor, R. Fergus, Y. LeCun, C. Bregler, Convolutional Learning of Spatio-
temporal Features, Springer Berlin Heidelberg, Berlin, Heidelberg, 2010,
nents, binarizing, and selecting training data for class balancing. Another pp. 140–153, https://doi.org/10.1007/978-3-642-15567-3_11. https://doi.org/10.
direction for future work is to evaluate the method of using multi-scale 1007/978-3-642-15567-3_11.
CNNs as an option for using anatomical information. [15] R. Bellman, Dynamic Programming, Courier Corporation, 2013.
[16] K. Kamnitsas, C. Ledig, V.F. Newcombe, J.P. Simpson, A.D. Kane, D.K. Menon,
There are some limitations in the current study. The images were
D. Rueckert, B. Glocker, Efficient multi-scale 3d CNN with fully connected CRF for
acquired at a single radiology department, and more images from accurate brain lesion segmentation, Med. Image Anal. 36 (2017) 61–78. http://doi.
different institutions would be needed to improve the generalizability. org/10.1016/j.media.2016.10.004. http://www.sciencedirect.com/science/
The smallest stones, that are highly likely to pass spontaneously and the article/pii/S1361841516301839.
[17] J. Jendeberg, H. Geijer, M. Alshamari, B. Cierzniak, M. Liden, Size matters: the
largest stones that are unlikely to be found in the distal ureter were width and location of a ureteral stone accurately predict the chance of spontaneous
excluded from the study. No scans without ureteral stones were included passage, Eur. Radiol. 27 (11) (2017) 4775–4785. Springer.
in the dataset. However, the main outcome measures - sensitivity and [18] G. Litjens, T. Kooi, B.E. Bejnordi, A.A.A. Setio, F. Ciompi, M. Ghafoorian,
J.A.W.M. van der Laak, B. van Ginneken, C.I. Sanchez, A survey on deep learning in
false positives rate - are not expected to be altered with negative controls. medical image analysis, Med. Image Anal. 42 (2017) 60–88, https://doi.org/
In conclusion, the present work describes a CNN method for detecting 10.1016/j.media.2017.07.005.
ureteral stones in thin slice CT. With further development, primary [19] B.J. Erickson, P. Korfiatis, Z. Akkus, T.L. Kline, Machine learning for medical
imaging, RadioGraphics 37 (2) (2017) 505–515, https://doi.org/10.1148/
reading of CTs by an automated detection algorithm is feasible. rg.2017160130 pMID: 28212054.
[20] M. Havaei, A. Davy, D. Warde-Farley, A. Biard, A. Courville, Y. Bengio, C. Pal, P.-
Conflicts of interest M. Jodoin, H. Larochelle, Brain tumor segmentation with deep neural networks,
Med. Image Anal. 35 (2017) 18–31. http://doi.org/10.1016/j.media.2016.05.004.
http://www.sciencedirect.com/science/article/pii/S1361841516300330.
None declared. [21] H.R. Roth, L. Lu, A. Farag, H.-C. Shin, J. Liu, E.B. Turkbey, R.M. Summers,
Deeporgan: multi-level deep convolutional networks for automated pancreas
segmentation, in: Medical Image Computing and Computer-assisted Intervention,
Acknowledgments
Springer, 2015, pp. 556–564.
[22] W. Li, F. Jia, Q. Hu, Automatic segmentation of liver tumor in CT images with deep
The authors would like to acknowledge Antai Llaquet for his contri- convolutional neural networks, J. Comput. Commun. 3 (11) (2015) 146–151,
butions. This work has been sponsored by Nyckelfonden (grant OLL- https://doi.org/10.4236/jcc.2015.311023.
[23] S.A. Pfister, A. Deckart, S. Laschke, S. Dellas, U. Otto, C. Buitrago, J. Roth,
597511) and by Vinnova under the project ”Interactive Deep Learning for W. Wiesner, G. Bongartz, T.C. Gasser, Unenhanced helical computed tomography vs
3D image analysis”. intravenous urography in patients with acute flank pain: accuracy and economic
impact in a randomized prospective trial, Eur. Radiol. 13 (11) (2003) 2513–2520,
https://doi.org/10.1007/s00330-003-1937-1.
References

[1] F.L. Coe, A. Evan, E. Worcester, Kidney stone disease, J. Clin. Invest. 115 (10)
(2005) 2598–2608.

160

You might also like