2.1.

Data Collection

We imaged cochleae in live mice using OCT, as described previously.^10,26 Briefly, mice were anesthetized with ketamine/xylazine, fixed to a head post to reduce motion artifacts during image acquisition, and underwent surgery to open the left middle ear bulla. We then imaged the cochlea without violating the otic capsule bone that surrounds it. All protocols were approved by the Institutional Animal Care and Use Committee at Stanford University.

Over the course of these experiments, we imaged the cochleae of 16 normal CBA/CaJ mice. To evaluate the effectiveness of our technique in classifying endolymphatic hydrops, five mice were exposed to a single blast pressure wave with a peak pressure of 130 kPa and four mice were exposed to band-passed white noise (8 to 16 kHz) at 100-dB sound pressure level for 2 h using previously described techniques.^27–29 We have found that mice develop endolymphatic hydrops after blast or noise exposure¹⁹ under these conditions. (A separate paper detailing the biological features of this phenomenon is currently in preparation.) The other seven mice served as the control group and received no blast or noise exposure.

We imaged each cochlea from multiple positions and depths at a fixed angle to capture hundreds of cross-sectional images per cochlea. The angle and depths of imaging were chosen to image the midmodiolar section of the cochlea [as in Figs. 1(b)–1(d)]. The purpose of this was to keep the angle of incidence consistent across experiments for visualizing Reissner’s membrane deformation. In three mice, some of the images were obtained using laser angle adjustments between 15 and $-15\mathrm{deg}$ from the midmodiolar plane to evaluate the influence of angle perturbations on displacement measurements. We found similar displacement measurements in images with and without angle adjustments, and included all data in our analysis. Our results were similar with and without inclusion of images at adjusted angles. In each frame, we focused our analysis on an $80\times 90\text{pixel}$ rectangle containing the apical turn of the cochlea, obtained by manually cropping the raw image. The optical resolution of the imaging system was $9.8\mu \mathrm{m}$ lateral and $15\mu \mathrm{m}$ axial measured in air.¹⁰ The higher refractive index in the cochlea (perilymph, 1.34)³⁰ improves the axial resolution; however, tissue-induced aberration and dispersion degrade the image quality. The sampling in the acquired images was $7.5\times 7.5\mu \mathrm{m}$ with interpolation done to obtain the correct aspect ratio.

In 10 mice, we also collected volumetric images along a $150\text{-}\mu \mathrm{m}$ length of basilar membrane inside the cochlea. We acquired 20 cross-sectional images at intervals of $7.5\mu \mathrm{m}$ and stacked the images to reconstruct a three-dimensional (3-D) image of the cochlea. These data were then used to quantify endolymph volume, since this is truly a direct indicator of the severity of endolymphatic hydrops.

2.2.

Ground-Truth Labeling of Images

Ground-truth labels of endolymphatic hydrops were determined visually by a trained observer (J.K.) without blinding during image acquisition and later reviewed by a second observer (G.S.L.) during image preparation. In all mice, multiple cross sections were obtained in the same cochlea, and the entire set of cross sections were viewed together to determine the ground-truth label of the cochlea and its cross sections. Of the 6391 images assessed, 3062 images were labeled as nonhydrops (controls) and 3329 images were labeled as endolymphatic hydrops. Images labeled as endolymphatic hydrops were obtained in blast- and noise-exposed mice, and images labeled as nonhydrops were obtained in control mice, with no ambiguous cases identified by the observers.

2.3.

Manual Measurement of Endolymph Volume

The scala media was manually segmented in each cross section of a volumetric image. The sum of the areas of the segmented regions was then multiplied by the interval distance between adjacent cross sections ($7.5\mu \mathrm{m}$) to calculate the endolymph volume. Manual segmentation was performed using Amira 5.4 software (Visage Imaging, San Diego, California).

2.4.

Binary Classification of Endolymphatic Hydrops Using Endolymph Volume

Prediction of endolymphatic hydrops from manually measured endolymph volume was performed using the binary classifier model

2.5.

Manual Measurement of Reissner’s Membrane Displacement

We assessed the distension of Reissner’s membrane by measuring the displacement of the membrane’s midpoint from its normal position, defined as a straight line between the attachment points of Reissner’s membrane at the spiral limbus medially and the spiral ligament laterally [Fig. 1(d)]. We refer to this metric as the perpendicular displacement of Reissner’s membrane, and to the attachment points of Reissner’s membrane as its endpoints. The sign of the displacement is positive if Reissner’s membrane bows away from the scala media (increased endolymph volume, i.e., endolymphatic hydrops) and negative if it bows toward the scala media (decreased endolymph volume).

Manual segmentation was considered the gold standard for measuring the perpendicular displacement of Reissner’s membrane. Manual segmentation was performed using the freehand drawing tool in the MATLAB^® Image Processing Toolbox (The Mathworks, Natick, Massachusetts). With this tool, Reissner’s membrane was interactively traced on the displayed image of the cochlea. The first, middle, and last points of the tracing were used to estimate the normal midpoint of Reissner’s membrane and calculate its perpendicular displacement as described.

2.6.

Computer-Aided Detection of Endolymphatic Hydrops

Our method of computer-aided detection involves three steps: (1) manual localization of the scala media of the apical cochlear turn in the original, uncropped OCT image, (2) automated measurement of Reissner’s membrane displacement using the cropped scala media, and (3) automated prediction of endolymphatic hydrops by binary classification, using displacement as the predictor variable. Thus, the entire process should be considered semiautomated.

2.6.1.

Manual localization of apical cochlear turn

Our method classifies OCT images in which the scala media occupies most of the image center and is oriented as shown in Figs. 1(c) and 1(d). To obtain images that meet these constraints, we acquired images in the preferred orientation during experiments and manually localized the scala media afterward. For manual localization, the user selected a point in the center of the scala media of the apical turn in the original, uncropped image using the computer mouse (Fig. 2, red asterisk). The image was then cropped to an $80\times 90\text{pixel}$ rectangle centered on the selected point. Beyond this step, the workflow for classification was fully automated.

Fig. 2

Automated workflow for assessing endolymphatic hydrops. For illustration, the raw image of a mouse cochlear cross section (left) is analyzed step-by-step. The first step, cropping, requires the user to click inside the scala media (red asterisk) in the raw image. A rectangular region is then cropped (dashed blue box). Here, the cropped dimensions are $111\text{pixels wide}\times 116\text{pixels high}$. The rest of the analysis is fully automated. The cropped image is segmented from the background. The mask of the scala media (green shaded area) is overlaid on the original cropped image to segment Reissner’s membrane (red line). Three points are identified on Reissner’s membrane (blue asterisks) to calculate the RM perpendicular displacement. Numbered steps are explained in the text. Scale bars are $100\mu \mathrm{m}$. SM, scala media and RM, Reissner’s membrane.

2.7.

Averaging Measurements

Information from multiple images of the same cochlea can be combined to improve the accuracy of displacement measurements. To accomplish this, we calculated the running average of measurements from five images. This was done by averaging the measurement in each image with the measurements from its four preceding images. Averaged measurements were not calculated for the first four images in each experiment. Information from multiple images included spatially adjacent cross sections from a volume image of the cochlea and temporally adjacent cross sections recorded at the same position during small adjustments to imaging position and depth of focus.

2.8.

Statistical Analysis

Statistics were performed in MATLAB^®. $p$ values were calculated using the unpaired student’s $t$-test. A critical alpha level of 0.05 was used to consider statistical significance. The Pearson’s correlation coefficient $R^2$ was calculated to assess the goodness of fit of linear regressions. Data points displayed as outliers were included in all analyses.

3.1.

Displacement of Reissner’s Membrane is Associated with Endolymphatic Hydrops Severity

First, we wanted to validate that the manual measurements of endolymph volume correlated with ground-truth labels as determined by two trained observers. Volumetric scans were collected from five control mice and five blast-exposed mice. Since endolymph volume changes dynamically after blast exposure, we serially repeated the scans three times in each mouse at 1-h intervals, thus collecting 30 volumetric scans in total. We then manually measured endolymph volume in each scan and correlated the measurements with the observer labels [Fig. 3(a)]. Based on these data, we trained a binary classifier, using a classification tree approach, and determined an endolymph volume threshold of $8.945\mathrm{nL}/150\mu \mathrm{m}$ to distinguish hydrops and nonhydrops. Using this threshold, these data were found to have a strong correlation ($p<0.001$, chi-square test) with the classification by trained observers, our accepted gold standard. The sensitivity and specificity of classification were 100%, demonstrating separation of hydrops and nonhydrops measurements at this threshold.

Fig. 3

Manual measurement of endolymph volume and Reissner’s membrane displacement. (a) Box plot of manually measured endolymph volume along a $150\text{-}\mu \mathrm{m}$ length of basilar membrane of cochlea using volumtric OCT, categorized by the presence or absence of endolymphatic hydrops as defined by the agreement of two trained observers viewing single cross sections. Examples of the maximum and minimum volume measurements in each exposure group are depicted. Scale bars are $100\mu \mathrm{m}$. (b) Scatter plot of manually measured displacement of Reissner’s membrane versus manually measured endolymph volume in the same 10 mice. The linear regression is shown as the red dashed line.

Next, we examined whether displacement of Reissner’s membrane measured from a cross-sectional image reflects the endolymph volume. We manually measured the perpendicular displacement of Reissner’s membrane in cross-sectional OCT images in the same 10 mice as above. A total of 308 cross-sectional images in control mice and 307 images in blast-exposed mice were analyzed, or an average of 20.5 cross-sectional images per volume scans. We observed a larger displacement of Reissner’s membrane in blast-exposed mice at all three time points compared with in control mice ($34.12\pm 11.45$ versus $5.26\pm 4.73\mu \mathrm{m}$, $\text{mean}\pm \mathrm{SD}$; unpaired $t$-test, $p<0.001$) [Fig. 3(b)]. Based on our data, the linear regression for endolymph volume as a function of Reissner’s membrane displacement was

3.2.

Robustness of Automated Segmentation of Reissner’s Membrane

The automated method for measuring Reissner’s membrane displacement relies on segmentation of Reissner’s membrane. The method returns a measurement of Reissner’s membrane displacement, if it successfully segments Reissner’s membrane, and returns no measurement (i.e., “NaN”) otherwise. Automated analysis could be performed correctly in every mouse. However, not all OCT images were analyzable.

Our dataset of 6391 images was split into a training dataset of 5287 training images from 13 mice (including the 10 mice described above; six control mice and seven blast- or noise-exposed mice) and test dataset of 1104 images from three mice (one control and two noise exposed) for developing the binary classifier for computer-aided detection of endolymphatic hydrops. We found that 2665 of the 5287 training images (50.4%) and 958 of the 1104 test images (86.8%) could be analyzed by the method. The variation in the percentage of analyzable images between the training and test datasets suggests that the image quality, which can vary for different experiments, is important for analyzability.

In particular, we found that the appearance of Reissner’s membrane as a continuous structure is important. The software often had difficulty analyzing images in which Reissner’s membrane appeared to have gaps (Fig. 4, red arrows). This would result in the scala media and scala vestibuli being segmented together as one connected component, which would prevent segmentation of the scala media’s left border and Reissner’s membrane. Thus, this approach to the automated segmentation of Reissner’s membrane may not work for any single given image in isolation. However, since OCT images are acquired as a continuous stream during an experiment, this method will work if at least some of the images are clear enough to permit automated analysis.

Fig. 4

Representative cross-sectional images of cochleae for which automated segmentation of Reissner’s membrane could not be performed. There is a hole-like discontinuity in Reissner’s membrane in the images (red arrows). The discontinuity makes it difficult to separate the scala media and scala vestibuli during thresholding, which is necessary for successful segmentation of Reissner’s membrane. For these images, the automated method returns no measurement of Reissner’s membrane displacement. Scale bars are $100\mu \mathrm{m}$.

3.3.

Automated Measurement of Reissner’s Membrane Displacement

To verify if automated measurements can replicate manual measurements of Reissner’s membrane displacement, we performed automated measurements in the same 10 mice for which manual measurements were obtained. We compared automated and manual measurements using 30 representative cross-sectional images, one image per mouse per time point. We observed a direct association between automated and manual measurements of Reissner’s membrane displacement (Pearson’s correlation coefficient: $R^2=0.87$, $p<0.001$) [Fig. 5(a)]. The best-fit linear regression was

Fig. 5

Automated measurement and classification of training data. (a) Manual versus automated measurements of Reissner’s membrane displacement in the same mice described in Fig. 3. Measurements were performed for 30 representative OCT cross sections. (b) Box plots of autodetected displacements in the training dataset, obtained in 13 mice (including the 10 mice described above), with and without averaging of five adjacent measurements. Control ($n=2149$ measurements) and hydrops ($n=1056$) without averaging; control ($n=2111$) and hydrops ($n=1012$) with averaging. Green line shows the decision threshold of $7.36\mu \mathrm{m}$, which maximized Youden’s index for training data classification. Red crosses indicate outliers, defined as values greater than the 75th percentile or less than the 25th percentile by more than 1.5 times the interquartile range. (c) ROC curves showing the sensitivity–specificity trade-offs at different decision thresholds for classifying endolymphatic hydrops (positives) and controls (negatives). $k=1$, no averaging; $k=5$, averaging measurements across five adjacent images. (d) Five OCT cross sections, obtained sequentially in time, and their automated displacement measurements. The last (right-most) image was misclassifed as a control using the nonaveraged measurement (red number) and correctly classified as endolymphatic hydrops using the averaged measurement. The scala media segmentation (green shaded area) and three coordinates for calculating displacement of Reissner’s membrane (blue asterisks) identified by the computer program are shown. Scale bar is $100\mu \mathrm{m}$.

We then compared the automatically measured displacements to the ground-truth labels of the trained observers. As expected, the displacement was higher in images of cochlea with endolymphatic hydrops versus without endolymphatic hydrops ($15.77\pm 10.2$ versus $2.12\pm 7.3\mu \mathrm{m}$, $\text{mean}\pm \mathrm{SD}$; unpaired $t$-test, $p<0.001$) [Fig. 5(b), left two columns]. To reduce the variability in the displacement data, we created a running average of the displacement from five sequential images. This approach maintained a similar difference in Reissner’s membrane displacement between the presence and absence of endolymphatic hydrops but reduced the SD ($15.65\pm 6.9$ versus $2.11\pm 5.3\mu \mathrm{m}$, $\text{mean}\pm \mathrm{SD}$; unpaired $t$-test, $p<0.001$) [Fig. 5(b), right two columns]. For example, in Fig. 5(d), the automatically measured displacement of Reissner’s membrane in the last image (on the farthest right) was $1.7\mu \mathrm{m}$ without averaging and $12.3\mu \mathrm{m}$ with averaging, closer to the values measured in other cross sections. These data demonstrate that computer-aided classification can reliably identify endolymphatic hydrops when spatial averaging is used to improve the precision of the measurement. The necessary spatial averaging is readily achieved since OCT images are collected continuously during our experiments; hence, there is always a stream of images to work with in real time.

3.4.

Computer-Aided Classification of Endolymphatic Hydrops

We next developed a binary classifier that uses computer-aided detection of Reissner’s membrane to measure membrane displacement and predict endolymphatic hydrops status. To train the binary classifier, we performed supervised learning on 3205 computer-aided measurements of displacement and their associated ground-truth labels obtained from images in the training dataset. These data represent 2149 cross-sectional OCT images of cochlea without endolymphatic hydrops and 1056 images of cochlea with endolymphatic hydrops, and included all analyzable images in the training dataset.

We classified endolymphatic hydrops using the binary classifier model

Fig. 6

Classification performance of test data. (a) Automated measurements of Reissner’s membrane displacement in the test dataset obtained in three additional mice (not used in the training data). This included one control mouse and two mice with noise-induced endolymphatic hydrops. Green line shows the decision threshold of $7.36\mu \mathrm{m}$. Control ($n=866$ measurements) and hydrops ($n=92$) without averaging; control ($n=864$) and hydrops ($n=90$) with averaging. (b) ROC curves for the test data, with and without averaging measurements (red line and blue line, respectively). (c) Three OCT cross-sectional images in the test dataset from a mouse with endolymphatic hydrops. Two of the images were misclassifed as controls and one was correctly classified as endolymphatic hydrops. Each row shows the same image in a different step of the image analysis workflow. The first column shows the automated segmentation of the scala media (green shaded area); the second column shows the automated segmentation of Reissner’s membrane (red line); and the third column shows the three coordinates (3 points; blue asterisks) used to automatically calculate Reissner’s membrane displacement. Scale bars are $100\mu \mathrm{m}$. (d) Confusion matrix of training data classification. (e) Confusion matrix of test data classification.

Next, we evaluated the trained classifier on the test dataset, which included data from three previously unused mice. This test dataset included 958 automatically measured displacements from 1104 test images and their ground-truth labels. This represented 866 images of cochlea without endolymphatic hydrops and 92 images of cochlea with endolymphatic hydrops, from one control mouse and two noise-exposed mice, respectively. We observed higher automatically measured displacements in images with endolymphatic hydrops compared with in images without endolymphatic hydrops, both without averaging ($3.77\pm 4.9$ versus $16.22\pm 10.9\mu \mathrm{m}$, $\text{mean}\pm \mathrm{SD}$; unpaired $t$-test, $p<0.001$) and with averaging of measurements ($16.21\pm 7.0$ versus $3.77\pm 3.1\mu \mathrm{m}$, $\text{mean}\pm \mathrm{SD}$; unpaired $t$-test, $p<0.001$) [Fig. 6(a)]. The ROC curve AUC for test data was 0.874 without averaging and 0.961 with averaging [Fig. 6(b)]. The classification of test data showed good sensitivity (0.8261) and specificity (0.8256) without averaging, and higher sensitivity (0.9111) and specificity (0.8773) with averaging such as classification of training data [Figs. 6(d)–6(e)]. Overall, our classifier showed a similar performance between the training and test datasets.

3.5.

Error Analysis

Classification of the test dataset resulted in 16 false negatives and 156 false positives. Images corresponding to two of the false negatives and one of the true positives are shown in Fig. 6(c). Each row shows an image at three different steps (shown in different columns) of computer-aided segmentation of Reissner’s membrane. From left to right, the columns show the scala media segmentation (green shaded area), Reissner’s membrane segmentation (red line), and autodetected coordinates (blue asterisks) for calculating Reissner’s membrane displacement.

We observed that misclassification frequently occurred because of errors in segmenting Reissner’s membrane. For example, the top row of Fig. 6(c) shows an image in which the upper end point of Reissner’s membrane was located too close to the midpoint of the membrane. This resulted in underestimation of the membrane’s displacement. (This image also shows an example of the correction to a sharp bend artifact discussed in step 6(a) of the automated workflow for segmenting Reissner’s membrane (Fig. 2). The initial segmentation of Reissner’s membrane [Fig. 6(c), red line, middle column] incorrectly includes the attachment area of the tectorial membrane to the spiral limbus; however, step 6(a) removes this portion of the segmentation before determining the end points [Fig. 6(c), blue asterisks, right column].)

Another example of misclassification due to a mistake in segmenting Reissner’s membrane is shown in the middle row of Fig. 6(c). In this image, the lower end point of Reissner’s membrane was identified up and to the left of its true location. This also resulted in underestimation of Reissner’s membrane displacement and caused misclassification of the image as not having endolymphatic hydrops (false negative). Clearly, improving the automated segmentation of Reissner’s membrane will be important for increasing the sensitivity and specificity of endolymphatic hydrops classification by this method.