Optical coherence microscopy as a novel, non-invasive method for the 4D live imaging of early mammalian embryos

Imaging of living cells based on traditional fluorescence and confocal laser scanning microscopy has delivered an enormous amount of information critical for understanding biological processes in single cells. However, the requirement for a high numerical aperture and fluorescent markers still limits researchers’ ability to visualize the cellular architecture without causing short- and long-term photodamage. Optical coherence microscopy (OCM) is a promising alternative that circumvents the technical limitations of fluorescence imaging techniques and provides unique access to fundamental aspects of early embryonic development, without the requirement for sample pre-processing or labeling. In the present paper, we utilized the internal motion of cytoplasm, as well as custom scanning and signal processing protocols, to effectively reduce the speckle noise typical for standard OCM and enable high-resolution intracellular time-lapse imaging. To test our imaging system we used mouse and pig oocytes and embryos and visualized them through fertilization and the first embryonic division, as well as at selected stages of oogenesis and preimplantation development. Because all morphological and morphokinetic properties recorded by OCM are believed to be biomarkers of oocyte/embryo quality, OCM may represent a new chapter in imaging-based preimplantation embryo diagnostics.

For the purposes of this study, we introduced a diversified time interval scanning protocol (DTIsp) that supports multiple time 102 intervals between the acquisitions of consecutive OCM measurements (see Supplementary Figure S1). DTIsp enables 103 researchers to perform a complex analysis of biological processes with different dynamics that occur in oocytes and embryos 104 at the expense of data oversampling. Nonetheless, the oversampled data can be averaged to reduce speckles and produce high-105 quality structural images or were transformed with a wide spectrum of mathematic operations. The latter included operations 106 such as difference, mean value, maximal value, minimal value or standard deviation and allowed us to attenuate or enhance 107 particular structures on final OCM images.

108
We designed DTIsp protocols to support up to four different time intervals at which each single point within the 3D volume 109 could be analyzed. Shortest time step, given by the CMOS camera repetition rate, defined the rate at which single axial scans 138 1) We first blur the 3D image using a Gaussian filter with 0.8

 
voxels to segment the cell body. Next, for each 2D layer at a given depth z (for each z coordinate) in the 3D frame, we calculate a separate threshold, tz , using Otsu's method 1 . To 140 make the segmentation more robust, tz is then averaged with the analogous thresholds obtained for the 50 nearest layers, i.e., 141 from 25 tz  to 25 tz  , resulting in the adjusted threshold tz . Next, the thresholds are used to classify the voxels in each 142 layer; the voxels in the zth layer with values greater than tz are assumed to represent the cell. Because some cell structures, 143 like pronuclei, are much darker than the cell body in OCM, their voxels may have lower values than tz and form 3D 'holes' 144 (cavities) in the thresholding outcome. We label all such holes as cell bodies to obtain a continuous 3D region of adjacent 145 labeled voxels that includes these structures. We then apply a morphological opening to remove the very small or narrow 146 parts of the cell body. This process may lead to several isolated objects, the largest of which (in terms of the volume) is 147 assumed to be the cell body.

148
When the male pronucleus is very close to the cell border, the adjacent cell membrane may not be clearly rendered, and the 149 above procedure may lead to an apparent cavity on the cell surface. We incorporate the information on the locations of 150 pronuclei in the previous frame (estimated using the method described below in 2), when available, to overcome this problem.

151
First, we identify the points located on the junction of three regions: the cell body detected in the current frame, the pronuclei 152 detected in the previous frame, and the background detected in the current frame. Next, we build a 3D convex hull on these 153 junction voxels, and label all points in the hull as belonging to the cell body. Finally, we fill with labels any cavities that might 154 have resulted from that process; thus, the result is guaranteed to form a continuous 3D region of adjacent labeled voxels.

155
2) Once the cell body is extracted from a frame, we detect the pronuclei or a spindle, depending on the phase of cell cycle 156 being analyzed. In both cases, the procedure works almost identically; therefore, we refer to both pronuclei and a spindle as For the frames that contain the spindle, the distribution of voxel values in the cell body is usually significantly different from 162 the distribution in the remaining frames, which required us to cap the threshold at a fixed value of 9.2. Additionally, because 163 the voxels near the cell surface are a bit darker than those throughout the cell interior, we slightly increased their values such 164 that they are less likely to be labeled as a pronucleus.

165
For each object identified in the previous frame, we approximate it with a small ellipsoid and scale it down. The voxels in the 166 current frame inside the ellipsoid then form 'seeds' for the watershed segmentation algorithm described below. First, we insert 167 all seed points, the information about the object's label (e.g., 'male pronuclei') and the lowest possible priority into a priority 168 queue. Next, we obtain the head element, i.e., the point with the lowest priority and its required label from the queue. If the 169 extracted point is not a seed and fewer than three of its neighbors either have the same label or are candidates to receive the 170 same label, then this point is only marked as a candidate to receive that label and is removed from the queue. In the other case, 171 the point is definitively labeled, and all of its neighbors are added to the priority queue. The priority of each added point is 172 calculated as the absolute difference between the value of the just labeled point and that of the added point. We ignore all 173 points with values above the ot threshold and points that have been already labeled. The process is repeated until the priority 174 queue is empty.

175
Once watershed segmentation is completed, we apply the morphological closing operation to the labeled voxels, fill any 3D 176 cavities in the labeled objects (in the same way as when segmenting the cell body in 1), and finally keep only the objects that 177 contain a center of any object detected in the previous frame. Each object segmented in this way is then approximated with 178 an ellipsoid; among others, that ellipsoid is used to seed the watershed segmentation in the next frame, as mentioned above.

179
Next, we verify whether the center of each ellipsoid lies inside the object detected in the previous image. If not, the 180 corresponding object is labeled as a non-pronucleus (implying that it may represent other cell structures).

181
As mandated by the underlying biological process, the pronuclei merge at a certain time point to form the spindle, which 182 requires special handling. We merge two pronuclei and label it as a spindle if (i) they have adjacent voxels and (ii) there is no 183 clear partition between them in terms of voxel values. We temporarily merge both adjacent pronuclei into one region and 184 apply the Multi Otsu method with two thresholds to verify the latter condition 2 . Next, we threshold the merged region with approximate the positions of the spindle in the previous and current frame with ellipsoids, and retrieve their longest axes, , we consider these values as a significant change in the 190 shape and/or volume of a spindle, and attempt to detect the objects again. Therefore, we apply a procedure based on the Multi

191
Otsu method 2 and watershed segmentation described above. If the procedure succeeds at detecting two new objects that 192 overlap with one spindle in the previous frame, nuclear division is detected.

193
3) The last step of automatic image segmentation is to detect the nucleoli inside the pronuclei or nuclei (in case of 2-cell 194 embryos). This procedure processes each pronucleus or nucleus (i.e., object) separately. First, we calculate the average voxel 195 value m of the given object, and set all voxels outside the object to m. Next, adaptive thresholding is used to detect the darker 196 areas, which are considered potential nucleoli. The threshold is calculated by convolving the image with a Gaussian filter and 197 subtracting a constant. Finally, we apply the morphological closing operation to merge the regions that are close to each other.

198
Each resulting isolated region is considered a separate nucleolus.

199
When carefully tuned, these methods successfully detected all required cellular structures in the sequences of raw OCT frames.

200
Only one sequence required additional manual marking of areas that should not be classified as part of the pronucleus.

201
However, only one frame in that sequence had to be marked in this way; the markers were then properly propagated to 202 consecutive frames and thus no other frames required manual marking.