Professional Documents
Culture Documents
3-D Human Sperm Flagellum Tracing in Low SNR Fluorescence Images
3-D Human Sperm Flagellum Tracing in Low SNR Fluorescence Images
3-D Human Sperm Flagellum Tracing in Low SNR Fluorescence Images
0278-0062 © 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
HERNANDEZ-HERRERA et al.: 3-D + t HUMAN SPERM FLAGELLUM TRACING 2237
Fig. 1. Algorithm overview. (a) A volume rendering of the 3D image stack (all sub-images have an identical scale bar); (b) contrast enhancement:
the original 3D image stack is pre-processed to improve the contrast; (c) a one-class classification algorithm is employed to enhance the flagellum
where the scale represents the probability that a voxel belongs to a tubular structure; (d) an iterative algorithm is employed to extract the flagellum’s
centerline; blue and green segments depict different iterations and red path depicts the last iteration; (e) extracted centerline for the 3D image stack
where the scale depicts the z-coordinate from the centerline; (f) flagellum’s centerline for different time points; each color identifies a different time
point. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)
back-tracing from terminal points. The main drawback of 2D+t images [23]–[27], extending here the method for 3D+t,
these approaches is the need to set the initial and terminal and it is used as the starting point of the centerline algorithm.
point, which usually produces shortcuts when tracking a large The algorithm searches for the best minimal path over a prede-
centerline if the two points are close. fined distance of approximately the maximum flagellum’s
Hernandez-Herrera et al. [14] recently proposed a one-class diameter (using the fast marching algorithm), just to move
classification and a machine learning algorithm (MESON) to a small distance from the starting point and then it stops. The
trace the centerline from tubular structures. Tubular structures cost to travel across points near the detected minimal path
were segmented using the classifier and the fast marching is forced to increase (in order to avoid detecting the same
algorithm allowed extraction of their centerline. An alternate minimal path), consecutively indexing the terminal point as the
iterative method (RIVULET2) to extract the centerline from starting one; this procedure is iterated to complete the whole
tubular structures was used by Liu et al. [20], also based flagellum’s centerline, avoiding detection of erroneous paths,
on the fast marching algorithm. Because these approaches specifically at the more curved parts. Our team has already
extract multi-branch structures by design, they can also be presented a preliminary work on sperm flagellum tracing [28].
used to extract single tubular structures. Indeed, multi-branch Compared to MESON and RIVULET2, our algorithm does
algorithms can extract correctly single tubular structures if the not rely on the segmentation of the tubular structures to
images have high contrast and high SNR. However, multi- trace the centerline, instead it relies on the enhanced stack
branch algorithms usually produce spurious branches from low to guide the tracing. Furthermore, those algorithms propagate
contrast images (as those in this work) when a single structure the fast marching until all the points in the segmentation
should be detected. had been visited, whereas our algorithm propagates the fast
Here we propose a tracing algorithm designed to extract marching until a pre-defined length has been reached, avoiding
single tubular structures from a 3D+t low contrast and noisy shortcuts, specifically at the most curved part of the tubular
image stack. We obtain the centerline from the tubular struc- structure. Finally, those algorithms back-trace from the local
tures using a machine learning approach combined with a maximums (there can be many of them) of the geodesic
minimal path algorithm (described in Sec. II C-D). The initial distance generated by the fast marching, instead our algorithm
and terminal point to propagate the fast marching are auto- only back-traces from the global maximum (a single point).
matically detected by the algorithm. The algorithm has been Therefore, our algorithm is better suited to trace the flagellum
designed to trace several single tubular structures avoiding since it traces a single branch, whereas the others may trace
the shortcuts produced by classical fast marching approaches. many branches.
First, the MESON [14] algorithm is applied to each 3D stack The main contributions of this paper are: (i) a novel method-
to enhance tubular structures. Then, the center of the sperm’s ology to trace the centerline from multiple single tubular
head is automatically detected as described in other works for structures with high curvature and low SNR, allowing to trace
2238 IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. 37, NO. 10, OCTOBER 2018
Laplacian. Hence, from the remaining voxels that were not Algorithm 3 Iterative Centerline Tracing
detected in the training set, some of them belong to the Input: 3D original image stack V , pre-processed 3D image
background and some others to the foreground. The one-class stack I , 3D enhanced image stack R f , iteration length l0
classifier is used to assign low values for those remaining and threshold T
voxels in the background while high values for the voxels Output: centerline tracing as a graph representation
in the foreground. Step 1: Detect starting point sk for tracing sperm k
2) Feature Extraction: Frangi et al. [33] demonstrated that Step 2: Compute fast marching algorithm starting at sk until
the eigenvalues (|λ1 | ≤ |λ2 | ≤ |λ3 |) of the Hessian matrix it reaches a length l0
provide local shape information of the 3D image stack. Specif- Step 3: Extract the centerline by back-propagating from the
ically, it can be shown that for a bright tubular structure λ2 , λ3 point pk that reached the length l0
are negative and have large absolute values and |λ1 | is low, Step 4: Increase the values of cost-function F( x ) for points
for a bright blob structure λ1 , λ2 and λ3 are negative and have inside a tube. This tube is constructed from the centerline
large absolute values, for a bright plate-like structure |λ1 |, |λ2 | detected in previous step and the radius of the flagellum
have a low value and λ3 is negative and has a large absolute Step 5: Compute the orientation and probability of the
value, and for a noisy structure (background) |λ1 |, |λ2 | and centerline to be inside the flagellum. If the probability
|λ3 | usually have low values. Hence, the eigenvalues are is smaller than 0.5 or orientation is larger than 110◦ stop
good features to differentiate a tubular structure from a noisy algorithm; otherwise return to Step 2 where starting point
structure. sk is set to pk .
3) One-Class Classifier Construction: For each scale σr
provided by the user, a one-class classifier is constructed (λ1
is not taken into account here). The one-class classifier is built
D. Iterative Centerline Tracing
from the distribution of the eigenvalues from the training set
Tσr by using a 2D histogram. The 2D histogram bins contain Algorithm 3 depicts the main steps of our iterative approach
the votes of the training set eigenvalues. If the feature vector for centerline tracing.
from a point in the 3D image stack is assigned to a bin with 1) Starting Point Detection: Our approach to track sperm is
high value, then the point should correspond to a background similar to previous algorithms [23], [34], [35] with the main
point. On the other hand, the feature vector from a point difference that the other algorithms were designed to track
inside a tubular structure should correspond to a bin with low the sperm head in 2D+t images, whereas our approach is
value in the 2D histogram since |λ2 | ≈ |λ3 |, |λ2 | >> 0 and developed for 3D+t images. The starting point sk used to
none or few vector in the training set must have high values.1 initialize Algorithm 3 at sperm k is selected at the center of
Therefore, the 2D histogram can be used to enhance tubular the sperm’s head (the algorithm can trace more than one sperm
structures. The 2D histogram is transformed to have values in per stack). Sperm heads are usually the brightest structures in
the range [0,1] and smoothed to have low transitions between the 3D image stack, this information is used to detect each
neighboring bins. Let’s call Dσr the final 2D histogram. Dσr seed point sk . The heads are segmented by applying a binary
can be used for classification purposes by: computing the threshold T . The threshold value should be low enough to
eigenvalues (λ2 and λ3 ) for a given position x in the 3D image allow segmentation of the voxels in the head but high enough
stack, calculating the bin position of the eigenvalues in Dσr to avoid segmentation of voxels in the flagellum. A connected
and selecting the value of Dσr in the assigned bin position. components algorithm is employed in the segmented stack to
Finally, the assigned value can be thresholded and everything identify each connected component Ck which corresponds to
above the threshold would be detected as belonging to the each sperm’s head. The starting point sk is computed for each
background, otherwise it would be detected as belonging to connected component as the centroid from the 3D coordinates
foreground. Instead of using Dσr for classification purposes, of the voxels in each connected component. To keep track of
here Dσr is used to enhance the foreground. sperm, an ID is assigned to each sk at the current time point (t)
4) Enhanced 3D Stack Construction: The scale σmax ( x ) used which depends on the IDs from the previous time (t − 1).
to compute the Laplacian is detected for each x in the 3D To this end, the nearest neighbor algorithm is used to do
image stack. Thus, the feature vector (eigenvalues λ2 and the assignment. Let rk be the seed points detected in the
λ3 ) is extracted and the assigned value on the 2D histogram previous time. Then, the starting point sk is assigned with
Dσmax is calculated for the feature vector. Let’s call this value the same track ID as sperm i from the previous time if the
R(x ), then according to our previous analyses R( x ) must have minimum Euclidean distance between sk and all the seed
high value (close to one) for points in the background but points in the previous time rk corresponds to sperm i (seed
low (close to zero) for the flagellum. Hence, the function point ri ). Our algorithm fails when two sperm are crossing
R f (
x ) = 1 − R(x ), is used to enhance the flagellum. Fig. 1(d) since the segmentation produces a single connected component
depicts an example of flagellum enhancement step. Note that instead of two and one head disappears. After the sperm cross,
the flagellum’s full structure can be visualized since low values the algorithm would detect two heads and the one that was not
are given in the background while high in the flagellum. detected will be assigned a new ID.
Fig. 2(a) depicts a 3D image stack with two nearby sperm.
1 The points in the training set correspond to the background of the 3D Fig. 2(b) depicts the sperm’s head segmentation in gray
image stack and its eigenvalues should be low (|λ2 | ≈ |λ3 | ≈ 0) while green points depict the centroid sk for each connected
2240 IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. 37, NO. 10, OCTOBER 2018
thus, among all the paths starting at sk and length l0 (there are
many paths satisfying this condition), we select the path with
minimum-cost (the one ending at pk ).
The cost function F( x ) plays an important role in the
detection of the minimum cost path since it guides the path.
The path will travel through low values of F( x ) instead of
higher. Hence, the function F( x ) should be constructed with
low values in the flagellum’s centerline and high values for
points distant from the centerline.
Fig. 2. Seed points detection. (a) volume rendering of the 3D image In our experiments the cost function F( x ) is defined as:
stack (all sub-images have an identical scale bar); (b) gray depicts
the segmented head, green depicts the detected seed point and red x ) = exp−60·G σa ∗[V (x )·R f (x )] ,
F( (2)
the detected point in the previous time point. (For interpretation of the
references to color in this figure legend, the reader is referred to the web where V is the raw image stack with values in the range [0, 1],
version of this article.)
R f is the enhanced flagellum, G σa is a Gaussian kernel and
“∗” represent the convolution, respectively. The function V
forces the minimum-cost path to travel across high intensity
component and red points depict the seed points rk from the
values of V (flagellum’s center) while the function R f discour-
previous Time Point (TP). The red points are used to assign
ages the path travelling outside the flagellum (R f assigns
the sperm’s ID to the green points using the nearest neighbor
low values to the background). The Gaussian smoothing G σa
algorithm.
is employed to reduce noise and have smooth cost values
2) Minimal-Cost Path: Our algorithm to trace the flagellum’s
between neighboring points, this creates smooth minimum cost
centerline is an iterative approach based on the minimal-cost paths, the σa was set to [1,1,1] equal to the minimum radius
path where the first iteration initializes at the starting points sk to be detected; a higher value may result in over-smoothing
and it stops when the paths reach a length l0 . The process is structures with radius smaller than σa . The value 60 was found
repeated and the terminal points from the previous paths are by parameter optimization using grid search and selecting the
used as the new positions for the starting points. This process value that stabilizes the highest average F-score using few
is repeated until the probability of the detected path to be in
stacks for training.
the flagellum is lower than 0.5 or there is a sharp change of The previous approach is designed to trace multiple sperm.
orientation in the path. The algorithm is initialized at each seed point sk corresponding
Given an initial point sk , a terminal point pk and a function to the center of each sperm’s head and the algorithm stops
F(x ) that determines the cost of traveling across point x, when all the paths C(sk , pk ) reached a length l0 .
the minimum-cost path to travel from sk to pk is given by
4) Increase Cost-Values for Already Visited Point: In order
a minimization problem
to discourage future iterations to travel across the already
pk detected paths C(sk , pk ), the cost function is increased for
T ( pk ) = min F(C(s))ds points close to C(sk , pk ). Recall that the flagellum resembles
C sk a tubular structure and the path C(sk , pk ) should travel across
the center of this tubular structure. Hence, a tube is constructed
where C(s) is a curve in R3 . Among all the possible paths to reduce the cost-function inside the tube. The tube is
to travel from sk to pk the path C(sk , pk ) with minimum constructed such that the starting point of the tube is the point
value is called the minimum-cost path. Given a point pk ∈ R3 , sk , the terminal point of the tube is pk and the center of the
the minimum-cost path C(sk , pk ) can be obtained by solving tube is given by the path C(sk , pk ), the radius of the tube is
the Eikonal equation which is a non-linear partial differential r0 and it is given by r0 = 2 · max{DT (C(sk , pk ))}, where
equation given by ∂T∂t(x ) = F(
x ), with T (sk ) = 0. The fast DT ( x ) is the distance transform [38] from the segmented
marching algorithm [36] is used to get an approximation to flagellum (obtained by applying a threshold of 0.5 to R f ( x ))
the solution T ( x ) of the Eikonal equation. and the radius r0 corresponds approximately to twice the
3) Extract the Centerline by Back-Propagating: Given a maximum radius of the flagellum in the detected centerline
starting point sk and a terminal point pk , the minimum cost C(sk , pk ). The cost function F( x ) is updated as:
path can be detected by backtracing from pk along the gradient
of T (x ) until sk is reached2 (see [37]). The fast marching ∞ if x belongs to any tube
F(x) =
algorithm is initialized at sk and the point pk is automatically F(x ) otherwise,
detected as
p the new cost function F( x ) discourages the minimal-cost path
to travel across points in the tube since the new cost value for
pk = argmin{ F(C(s))ds such that C = l0 },
p sk traveling across those values is infinitely expensive (∞).
Fig. 1(d) depicts the different iterations of the algorithm for
2 Get all the neighboors of p a 3D image stack with one sperm, the first iteration starts at the
k and select the neighboor c0 with minimum
value T (x ). Get all the neighboors of c0 and select the neighboor c1 with sperm’s head and is depicted by a green path, the second by
minimum value T ( x ). Repeat until sk is reached a blue path, the third by a green path and so on. The red path
HERNANDEZ-HERRERA et al.: 3-D + t HUMAN SPERM FLAGELLUM TRACING 2241
Fig. 5. Qualitative results. (a) Volume rendering of a 3D stack with multiple sperm (all sub-images have an identical scale bar). (b) trace for each
sperm, colormap depicts the value of the extracted flagellum’s z-coordinate. (For interpretation of the references to color in this figure legend,
the reader is referred to the web version of this article.)
sperm’s head and propagating the fast marching algorithm for MESON algorithm as a blue line. Note that the algorithm
all these points. This allows the algorithm to avoid jumping stops prematurely for Calcein DS1 TP 50 (Fig. 6(l)) while
to another sperm. An even more difficult task is to trace it produces extra branches for Fluo-4 DS1 TP 28 (Fig. 6(k)).
the flagellum when two sperm are in close proximity and This algorithm ignores the a-priori knowledge that there is
the flagellum are touching or crossing, our current approach only one tubular structure to trace. Fig. 6(m-o) depicts the
cannot handle this task. centerline extracted by the RIVULET2 algorithm as a red
Fig. 6 depicts qualitatively results againts state-of-the- line. Note that RIVULET2 creates extra branches for the three
art tubular structures tracing algorithms (MESON [14] and stacks and it stops prematurely for Fluo-4 DS3 TP 6 and Fluo-
RIVULET2 [20]) where green, yellow, blue and red lines 4 DS1 TP 28 (Fig. 6(m,n)). Multi-branch algorithms such
represent the semi-manual trace, our trace, MESON trace as MESON and RIVULET2 can be modified to extract a
and RIVULET trace, respectively. An exhaustive grid search single structure by selecting the largest branch. However, this
for the best parameter configuration at each time point was approach may fail since there are time points where the largest
performed for the MESON and RIVULET2 algorithms. branch does not correspond to the flagellum’s centerline.
Fig. 6(a-c) depicts the volume rendering for three different Furthermore, the required time to perform this modification
time points (Fluo-4 DS3 TP 6, Fluo-4 DS1 TP 28 and would be even more expensive than manual trace since it
Calcein DS1 TP 50) from three 3D stacks with low, high and requires to do a grid search for parameter configuration (this
medium curvature. Fig. 6(a,b) corresponds to sperm loaded step can produce a lot of traces depending on the number of
with Fluo-4 while Fig. 6(c) corresponds to a sperm loaded parameters, the interval limits and step size), then manually
with Calcein. Note that the images corresponding to sperm select the best centerline among all the traces and finally
loaded with Fluo-4 have more noise and have less contrast than selecting the best branch. Indeed, the best traces depicted
the sperm with Calcein. The head is the brightest structure of in Fig. 6 for the MESON and RIVULET2 algorithms have
the 3D stacks and the flagellum is very difficult to visualize different parameter setting for each time point.
since it has very low contrast. Fig. 6(d-f) depicts the semi-
manual trace for each time point as a green line. The semi-
manual trace for each time point was extracted using the B. Quantitative Results
plugin Simple Neurite Tracer [39] from the open source image In this section, quantitative results are presented for
processing application FIJI [40]. To this end, the user selected six out of eight 3D+t image stacks. Calcein DS3 and
an initial point at the sperm head and another terminal point Calcein DS4 were disregarded from this analysis
in the flagellum. Then, the algorithm extracts the minimum- because the semi-manual trace was unfeasible for these
cost path between these two points. The user selected another large 3D+t image stacks (used only for qualitative
point in the flagellum and the minimum path is detected. The analysis). The metrics Precision (P), Recall (R), and
user repeated this process several times (to avoid shortcuts) the F-score (F) are employed to measure the quality
until the tail of the flagellum is reached. Then, the user has of the reconstruction, P = SC /(SC + Sextra ), R =
to proceed with the next time point. Fig. 6(g-i) depicts the SC /(SC + Smiss ), F = 2(P · R)/(P + R), where SC is
centerline traced by our approach as a yellow line. Note that the total length of the automatic trace that was correctly
the trace highly resembles the semi-manual trace in the three traced (true positive), Smiss and Sextra are the total length of
time points. Fig. 6(j-l) depicts the centerline traced by the the missing (false negative) and extra segments (false positive)
2244 IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. 37, NO. 10, OCTOBER 2018
Fig. 6. (a-c) Volume rendering from 3D stacks, the first two were treated with Fluo-4 while the third was treated with calcein (each column has an
identical scale bar); (d-f) semi-manual trace for each 3D stack; (g-i) centerline traced by our iterative algorithm; (j-l) centerline traced by MESON
algorithm; (m-o) centerline traced by RIVULET2 algorithm. (For interpretation of the references to color in this figure legend, the reader is referred
to the web version of this article.)
of the automatic trace, respectively. The values for each metric In our experiments, C was set to 7.5 voxels since this value
are in the range [0,1] with a low value corresponding to is the maximum flagellum’s diameter in our dataset. The last
a wrong tracing while high values correspond to good metric is Aver age − Di splacement − Err or (ADE) which
tracings. A point pc in the trace is classified correctly if measures the average distance error between the manual
there is a point pm in the semi-manual trace such that trace and the automatic trace. Note that Precision penalizes
pc − pm < C, where C is a constant defined by the user. the automatic trace if it deviates (it has extra segments)
HERNANDEZ-HERRERA et al.: 3-D + t HUMAN SPERM FLAGELLUM TRACING 2245
performed the worst for the 3D+t image stacks and therefore previous algorithms able to extract the centerline from this
it is discouraged to use it to trace a single branch struc- type of fluorescent images. Other algorithms able to extract
ture. Pre-processing the 3D image stack using the Frangi the centerline from tubular structures have been designed to
vesselness measure increases the average metric values for the extract more than one branch, however there is a prior knowl-
RIVULET2, however it still has low metric values indicating edge that the flagellum is a single structure. The quantitative
incorrect tracings. Comparing the average value between our and qualitative results show that the multi-branch algorithms
approach and the two variations of our algorithm (FLE and MESON and RIVULET2 produce poor results when tracing
FLR), we note that FLR has the lowest F-score (0.92) value a single structure, whereas our algorithm is producing similar
indicating that it performed the worst for the 3D+t image results as a manual tracing. The qualitative results using a
stacks and therefore it is discouraged to use the raw data to total of 564 semi-manual traces highly discourage using the
guide the minimal-cost path. FLE highly increases the metrics raw data to guide the minimal-cost path (Table I, method
values, it has an average F-score of 0.94, hence adding the C) and they encourage the use of the enhanced image to
enhanced image to guide the minimal-cost path is highly keep the trace inside the flagellum (Table I, method B).
recommended since it results in better tracing than using the Furthermore, the iterative approach increases the accuracy of
raw image. Finally, adding the iterative approach (our algo- the tracing (Table I, method A), specifically at the more curved
rithm) also increases the accuracy, it has an average F-score parts of the flagellum. The proposed method is capable of
of 0.97, thus for 3D+t image stacks with flagella beating solving a major problem related with the analysis of the 3D
with high curvature is recommended to included an iterative motility of sperm cells in fluorescence images with very low
approach. Finally, our algorithm has an average ADE value intensity and low SNR.
of 2.28, thus, indicating that the average distance between
the automatic trace and the semi-manual trace is very near ACKNOWLEDGMENT
to 2 voxels which suggest that the traces are very close to The author would like to thank P. Torres and E. Mata for
each other. helpful assistance with biological procedures and A. Bribiesca
The running times to trace the 564 stacks are: 1 hour (h) for computational support.
8 minutes (min) for our algorithm, 1 h 1 min for FLE,
12 min for FLR, 131 h 32 min for MESON, 12 h 26 min R EFERENCES
for RIVULET2 and 16 h for FRANGI+RIVULET. MESON [1] A. Darszon, T. Nishigaki, C. Beltran, and C. L. Treviño, “Calcium
required the highest reconstruction time because it needs to channels in the development, maturation, and function of spermatozoa,”
tune three parameters, hence a grid search for the best trace Physiol. Rev., vol. 91, no. 4, pp. 1305–1355, 2011.
[2] G. Corkidi et al., “Are there intracellular Ca2+ oscillations correlated
requires an excessive amount of parameter combination (847). with flagellar beating in human sperm? A three vs. two-dimensional
The lowest time was achieved by FLR with 12 min since analysis,” MHR, Basic Sci. Reproductive Med., vol. 23, no. 9,
it only requires one run of the fast marching algorithm to pp. 583–593, 2017.
[3] D. Oriola, H. Gadêlha, and J. Casademunt, “Nonlinear amplitude
obtain the best trace for each 3D stack. Note that FLE is dynamics in flagellar beating,” Roy. Soc. Open Sci., vol. 4, no. 3,
approximately 5 time slower than FLR. This is mainly because p. 160698, 2017.
the construction of the enhanced image stack R f ( x i ) is a [4] S. Ishijima, “Digital image analysis of flagellar beating and microtubule
sliding of activated and hyperactivated sperm flagella,” Soc. Reproduc-
time consuming task due to the calculation of the feature tion Fertility Suppl., vol. 65, pp. 327–330, 2007.
vector (eigenvalues of the Hessian matrix). The most time [5] A. Bukatin, I. Kukhtevich, N. Stoop, J. Dunkel, and V. Kantsler,
consuming task of our algorithm is the construction of the “Bimodal rheotactic behavior reflects flagellar beat asymmetry in
human sperm cells,” Proc. Nat. Acad. Sci. USA, vol. 112, no. 52,
enhanced image stack representing 67% of the total running pp. 15904–15909, 2015.
time (Sec. II-C, since it required the calculation of the Lapla- [6] T. Hyakutake, H. Suzuki, and S. Yamamoto, “Effect of non-Newtonian
cian, feature vector and one-class classifier). Algorithm 3 has fluid properties on bovine sperm motility,” J. Biomechanics, vol. 48,
no. 12, pp. 2941–2947, 2015.
to be re-run for traces that stopped prematurely, on average this [7] G. L. Takei, M. Fujinoki, K. Yoshida, and S. Ishijima, “Regulatory
value was 15% of the total traces. The maximum percentage mechanisms of sperm flagellar motility by metachronal and synchronous
was for Fluo-4 DS1 that required to re-run the algorithm 3 sliding of doublet microtubules,” MHR, Basic Sci. Reproductive Med.,
vol. 23, no. 12, pp. 817–826, 2017.
in 25 traces from a total of 73 traces (34%), whereas the [8] D. Lesage, E. D. Angelini, I. Bloch, and G. Funka-Lea, “A review
minimum percentage was for Calcein DS1 which required of 3D vessel lumen segmentation techniques: Models, features and
to re-run the Algorithm 3 only 6 traces from a total of 98 extraction schemes,” Med. Image Anal., vol. 13, no. 6, pp. 819–845,
Aug. 2009.
traces (6%). The total of traces to be re-run depends on the [9] E. Smistad, T. L. Falch, M. Bozorgi, A. C. Elster, and F. Lindseth,
quality of the stacks; if there are large gaps within them, “Medical image segmentation on GPUs—A comprehensive review,”
then many will stop prematurely. This can be usually fixed Med. Image Anal., vol. 20, no. 1, pp. 1–18, 2015.
[10] L. Acciai, P. Soda, and G. Iannello, “Automated neuron tracing methods:
by increasing the initial value l0 . However, increasing l0 to An updated account,” Neuroinformatics, vol. 14, no. 4, pp. 353–367,
a large value would induce the algorithm to miss the most 2016.
curved parts of the flagellum. [11] C. Becker, R. Rigamonti, V. Lepetit, and P. Fua, “Supervised feature
learning for curvilinear structure segmentation,” in Proc. Int. Conf. Med.
Image Comput. Comput.-Assist. Intervent. Berlin, Germany: Springer,
IV. C ONCLUSIONS 2013, pp. 526–533.
An iterative algorithm for the tracing of the sperm’s [12] Y. Zheng, M. Loziczonek, B. Georgescu, S. K. Zhou, F. Vega-Higuera,
and D. Comaniciu, “Machine learning based vesselness measurement
flagellum from a 3D+t image stack of very low SNR fluo- for coronary artery segmentation in cardiac CT volumes,” Proc. SPIE,
rescence based images has been developed. There are no vol. 7962, p. 79621K, Mar. 2011.
HERNANDEZ-HERRERA et al.: 3-D + t HUMAN SPERM FLAGELLUM TRACING 2247
[13] A. Sironi, V. Lepetit, and P. Fua, “Multiscale centerline detection by [27] X. Zhou, L. Ma, Y. Shang, M. Xu, X. Fu, and H. Ding, “Hybrid
learning a scale-space distance transform,” in Proc. IEEE Conf. Comput. generative-discriminative learning for online tracking of sperm cell,”
Vis. Pattern Recognit., Jun. 2014, pp. 2697–2704. Neurocomputing, vol. 208, pp. 218–224, Oct. 2016.
[14] P. Hernandez-Herrera, M. Papadakis, and I. A. Kakadiaris, “Multi-scale [28] P. Hernandez-Herrera, F. Montoya, J. M. Rendón, A. Darszon, and
segmentation of neurons based on one-class classification,” J. Neurosci. G. Corkidi, “Sperm flagellum center-line tracing in fluorescence 3D+t
Methods, vol. 266, pp. 94–106, Jun. 2016. low SNR stacks using an iterative minimal path method,” in Proc.
[15] A. Santamaría-Pang, P. Hernandez-Herrera, M. Papadakis, P. Saggau, Int. Conf. Image Anal. Recognit. Cham, Switzerland: Springer, 2017,
and I. A. Kakadiaris, “Automatic morphological reconstruction of pp. 437–445.
neurons from multiphoton and confocal microscopy images using [29] G. Corkidi, B. Taboada, C. D. Wood, A. Guerrero, and A. Darszon,
3D tubular models,” Neuroinformatics, vol. 13, no. 3, pp. 297–320, “Tracking sperm in three-dimensions,” Biochem. Biophys. Res.
2015. Commun., vol. 373, no. 1, pp. 125–129, 2008.
[16] D. Jiménez, D. Labate, I. A. Kakadiaris, and M. Papadakis, “Improved [30] F. Silva-Villalobos, J. A. Pimentel, A. Darszon, and G. Corkidi,
automatic centerline tracing for dendritic and axonal structures,” “Imaging of the 3D dynamics of flagellar beating in human sperm,”
Neuroinformatics, vol. 13, no. 2, pp. 227–244, 2015. in Proc. 36th Annu. Int. Conf. IEEE Eng. Med. Biol. Soc., 2014,
[17] H. Xiao and H. Peng, “APP2: Automatic tracing of 3D neuron pp. 190–193.
morphology based on hierarchical pruning of a gray-weighted image [31] N. Sharma, S. Saurav, S. Singh, R. Saini, and A. K. Saini,
distance-tree,” Bioinformatics, vol. 29, no. 11, pp. 1448–1454, “A comparative analysis of various image enhancement techniques
2013. for facial images,” in Proc. IEEE Int. Conf. Adv. Comput., Commun.
[18] M.-P. Garcia et al., “Coronary vein extraction in MSCT volumes using Inform. (ICACCI), Aug. 2015, pp. 2279–2284.
minimum cost path and geometrical moments,” IEEE J. Biomed. Health [32] S. Bedi and R. Khandelwal, “Various image enhancement techniques—
Inform., vol. 17, no. 2, pp. 336–345, Mar. 2013. A critical review,” Int. J. Adv. Res. Comput. Commun. Eng., vol. 2, no. 3,
[19] S. Liu, D. Zhang, S. Liu, D. Feng, H. Peng, and W. Cai, “Rivulet: pp. 1605–1609, 2013.
3D neuron morphology tracing with iterative back-tracking,” Neuroin- [33] A. F. Frangi, W. J. Niessen, K. L. Vincken, and M. A. Viergever,
formatics, vol. 14, no. 4, pp. 387–401, 2016. “Multiscale vessel enhancement filtering,” in Proc. Int. Conf. Med. Image
[20] S. Liu, D. Zhang, Y. Song, H. Peng, and W. Cai, “Auto- Comput. Comput.-Assist. Intervent. Berlin, Germany: Springer, 1998,
mated 3D neuron tracing with precise branch erasing and confi- pp. 130–137.
dence controlled back-tracking,” IEEE Trans. Med. Imag., to be [34] J. Liu, C. Leung, Z. Lu, and Y. Sun, “Quantitative analysis of locomotive
published. behavior of human sperm head and tail,” IEEE Trans. Biomed. Eng.,
[21] Y. Chen et al., “Curve-like structure extraction using minimal path vol. 60, no. 2, pp. 390–396, Feb. 2013.
propagation with backtracking,” IEEE Trans. Image Process., vol. 25, [35] S. T. Mortimer, “CASA—Practical aspects,” J. Androl., vol. 21, no. 4,
no. 2, pp. 988–1003, Feb. 2016. pp. 515–524, 2000.
[22] S. Basu and D. Racoceanu, “Reconstructing neuronal morphology from [36] J. A. Sethian, Level Set Methods and Fast Marching Methods Evolving
microscopy stacks using fast marching,” in Proc. IEEE Int. Conf. Image Interfaces in Computational Geometry, Fluid Mechanics, Computer
Process. (ICIP), Oct. 2014, pp. 3597–3601. Vision, and Materials Science, vol. 3. Cambridge, U.K.: Cambridge
[23] L. F. Urbano, P. Masson, M. VerMilyea, and M. Kam, “Auto- Univ. Press, 1999.
matic tracking and motility analysis of human sperm in time-lapse [37] M. S. Hassouna, A. A. Farag, and R. Falk, “Differential fly-throughs
images,” IEEE Trans. Med. Imag., vol. 36, no. 3, pp. 792–801, (DFT): A general framework for computing flight paths,” in Proc. Int.
Mar. 2017. Conf. Med. Image Comput. Comput.-Assist. Intervent. Berlin, Germany:
[24] Y. Imani, N. Teyfouri, M. R. Ahmadzadeh, and M. Golabbakhsh, “A new Springer, 2005, pp. 654–661.
method for multiple sperm cells tracking,” J. Med. Signals Sensors, [38] Y. Mishchenko, “A fast algorithm for computation of discrete Euclidean
vol. 4, no. 1, pp. 35–42, 2014. distance transform in three or more dimensions on vector processing
[25] O. Beya, M. Hittawe, D. Sidibé, and F. Mériaudeau, “Automatic detec- architectures,” Signal, Image Video Process., vol. 9, no. 1, pp. 19–27,
tion and tracking of animal sperm cells in microscopy images,” in Proc. 2015.
IEEE 11th Int. Conf. Signal-Image Technol. Internet-Based Syst. (SITIS), [39] M. H. Longair, D. A. Baker, and J. D. Armstrong, “Simple neurite tracer:
Nov. 2015, pp. 155–159. Open source software for reconstruction, visualization and analysis of
[26] H.-F. Yang, X. Descombes, S. Prigent, G. Malandain, X. Druart, neuronal processes,” Bioinformatics, vol. 27, no. 17, pp. 2453–2454,
and F. Plouraboué, “Head tracking and flagellum tracing for sperm 2011.
motility analysis,” in Proc. IEEE 11th Int. Symp. Biomed. Imag. (ISBI), [40] J. Schindelin et al., “Fiji: An open-source platform for biological-image
Apr./May 2014, pp. 310–313. analysis,” Nature Methods, vol. 9, no. 7, pp. 676–682, Jul. 2012.