Seedling fibroid characterization using optical coherence tomography

To the authors’ knowledge , this work presents the first OCT characterization of fibroids and normal tissues. In our work, we have quantitatively distinguished OCT volumes of normal and fibroid tissues within the human myometrium using optical, texture, statistical, and frequency-based features. We also characterized the collagen architecture using gradient-informed fiber orientation analysis. In this study, we extracted optical, texture, statistical, and frequency-based features in both the B-scan and en face directions. The fibroid is a distinct structure within the en face dimensions, and the projections capture depth-based changes and structural differences between the fibroid and normal regions. In both B-scan and en face dimensions, we found that fibroids exhibit lower kurtosis. In B-scans, energy could highlight the increased number of bright pixels and greater variation in pixel intensity, thus reflecting greater collagen disorder within the fibroid. Within the en face dimension, the fibroids had higher variance and lower kurtosis. This indicated that the pixel intensities of the normal specimens were more uniformly distributed. The attenuation coefficient does not significantly discriminate between the two classes. Light attenuation within fibroids can be inconsistent due to heterogeneous collagen content, leading to inconsistent trends. We also extracted statistical features from the distribution of gradients within en face sum-projection windows. The distribution ratio consistently distinguishes fibroid and normal specimens across depth. Fibroids showcase higher distribution ratios, indicating a more spread distribution. This finding aligns with those found from the texture, statistical, and frequency features. Increased heterogeneous collagen content within the fibroid creates greater variations in pixel intensities compared to well-aligned tissue, leading to greater variations in gradients within a window. The features extracted in this study are indirect indicators of tissue organization. The statistical analysis shown in Fig. 2 compares statistics extracted from intensity distributions of patches from en face projections. Since collagen is highly scattering in OCT, and the collagen fiber architecture within the fibroids is denser and disorganized [ 4 ], the statistics derived from the intensity distributions reflect this. The circular analysis completed in Fig. 4 is derived directly from the collagen fibers within the OCT images. The features shown in Fig. 2 could serve as indicators before collagen fiber analysis. Pairing this analysis with a higher-resolution modality such as SHG [ 32 ] could further elucidate collagen architecture. We did not find that any particular patients produced outliers in any analysis. The outliers in Figs. 2 , 3 , and 4J are from different patients, intramural fibroids, and various regions. Further characterizing seedling fibroids is important to further understanding how they will impact treatment plans. We did not observe differences between the fibroid specimens and those classified as other pathologies. Some of these “other” specimens contained fibrosis, but they could not be considered a fibroid because the fibrosis was not encapsulated. Additionally, the myometrium surrounding some of the fibroids appeared distorted, possibly due to adenomyosis. There are limitations to our study. One limitation is the size of the dataset. The diversity of patient demographics, disease history, and disease presentation impacts the interpretation of our results. Our dataset contained 11 patients, and some patients contributed more tissue than others. Additionally, factors such as gravidity, parity, hormonal status, and menstrual cycle stage varied among the patients and may affect the tissue presentation. The presence of other diseases, such as adenomyosis and fibrosis, also affected the myometrial structure within areas without a fibroid. Finally, other modalities, such as polarization-sensitive (PS)-OCT, can directly detect increased collagen content due to its birefringence [ 64 , 65 ] or, like SHG [ 32 ], visualize collagen fibers. In future work, a larger dataset will be acquired to address heterogeneity within the specimens and patient population. We will also use aspects of gynecologic history as biomarkers to gain additional insights into the structure and further characterize the myometrium. For example, Patients 8 and 9 were the only patients who took progestin as hormone therapy, and Patients 3 and 8 were never pregnant and did not have children. We could also further investigate the impact of location and fibroid type. Most of the fibroids in our study were intramural. The literature has established that fibroid presentation can differ within patients [ 1 ]. We observe the impact of disease and fibroid heterogeneity in our data, as seen in Figs. 2 and 3 . Additionally, due to OCT’s limited penetration depth and imaging range, future in vivo applications include a forward-viewing catheter integrated into surgical or therapeutic probes. Optical methods for screening may require methods with increased penetration depth, such as NIRS and photoacoustics. In conclusion, we demonstrated the first characterization of seedling fibroids using OCT-derived features and collagen fiber analysis. This work demonstrates the potential of OCT to elucidate the structure of seedling fibroids.

A total of 42 specimens were obtained from 12 patients post-hysterectomy in areas without visible pathology. Specimens and patient information were collected under an approved IRB protocol by Columbia University. Specimens were acquired from the fundus, anterior, posterior, and unspecified regions of the uterus. After non-destructive mechanical testing [ 51 , 52 ], the samples were frozen and stored at -80° C. The specimens were transferred to a -5° C freezer before imaging. The following demographics and gynecological history were recorded for each patient: age, race, ethnicity, gravidity, parity, estimated menstrual cycle stage, and hormonal therapy status. On the morning of imaging, following the protocol of previous studies [ 51 , 52 ], the samples were thawed at room temperature in phosphate-buffered saline for at least 20 minutes. The samples were imaged at room temperature. 37 samples from 11 patients were analyzed in this work. The specimens were imaged using a Thorlabs Telesto I OCT system with a center wavelength of 1325 nm and a bandwidth of 150 nm, an axial resolution of 6.5 µm, and a lateral resolution of 15 µm. The imaging range was 2.51 mm in air. Overlapping volumes of dimensions 5 mm × 5 mm × 2.51 mm were obtained. The specimens were flattened using a custom-built sample holder to enhance image quality. The OCT scan lens was inverted, and the samples were placed on quartz glass at an angle of 4° clockwise to reduce surface reflections. The reflection from the top of the quartz class results in a hyperintense line (observed in Figs. 1 and S3). The OCT data were registered using an algorithm previously described [ 53 , 54 ]. A fibroid specimen from Patient 8. Fibroids are highlighted in yellow. (A) A stitched, white light image obtained during OCT imaging. (B) Masson’s trichrome histology of the specimen. (C) An unstitched en face slice from the highlighted volume in A. (D) An unstitched standard deviation projection of the highlighted region in Panel A. (E, F, G) Sample stitched B-scans obtained from indicated locations in Panel C. We highlight the hyperintense lines, which are reflections from the first surface of the quartz glass. Fiber alignments differ between fibroid and adjacent myometrium. The scale bars represent 1 mm. After imaging, the specimens were placed in 10% formalin for at least 24 hours, then transferred to 70% ethanol, and, finally, sent to the Columbia University Pathology Core for histological processing. The slides were digitized at 20× with a Leica 516 Aperio AT2 slide scanner. The samples were stained using Masson’s trichrome staining to emphasize collagen. Histological analysis was conducted by a board-certified gynecologic pathologist (XC), blinded to the OCT images. The samples were labeled based on the clinical pathology summary. Table  1 summarizes patient and specimen information. In our analysis, we label the specimens “fibroid,” “normal,” and “other pathologies.” All of our fibroid specimens were seedling fibroids, i.e., those less than 1 cm in diameter [ 21 ]. Normal specimens contained adjacent myometrium and control (distant myometrium) tissue. The data labeled other pathologies contain adenomyotic and fibrotic tissues. 12 of our 37 specimens do not contain fibroid tissue. We performed fiber orientation analysis [ 55 , 56 ] to investigate the collagen architecture of these specimens. The dominant angle within a region, W, was determined by analyzing pixel-wise gradients. The fiber direction, θ , and concentration, b , were determined by fitting the 2D von Mises distribution (Eqn. ( 1 )) to the data [ 34 ]. I 0 (b) is a modified Bessel function of the first kind. Circular dispersion was derived after analyzing collagen architecture in sum projections. These projections had dimensions of 5 mm × 5 mm × ∼15/∼30/∼60 µm (1024 × 1024 × 4/8/16 pixels 3 ). In our study, we used window sizes of 396 × 396 µm 2 (81 × 81 pixels 2 ). We selected this window size based on work that characterized the collagen fiber organization within the cervix [ 34 ]. Windows with a mean intensity below a specified threshold were dropped because we used automated surface detection (see Fig. S1), and as a result, some regions within the en face projection lacked tissue and could not be reliably described. We extracted the circular mean, circular variance, circular standard deviation, circular skewness, and circular kurtosis using a MATLAB package [ 57 ]. We also extracted and reported the distribution ratio, the maximum value of the distribution divided by its standard deviation. (1) P ( x ) = e x p ( b c o s ( x − θ ) ) 2 π I 0 ( b ) We selected patch sizes for en face analysis based on previous work examining the collagen fiber organization within OCT en face images in the cervix [ 34 ]. Non-overlapping patches were extracted from unstitched standard deviation (SD) projections and sum en face projections. The extracted patches had depths of ∼15 µm, ∼30 µm, and ∼60 µm (4 pixels, 8 pixels, and 16 pixels), and en face dimensions of 469 × 469 µm 2 , 625 × 625 µm 2 , 781 × 781 µm 2 , 938 × 938 µm 2 , and 1250 × 1250 µm 2 (96 × 96 pixels 2 , 128 × 128 pixels 2 , 160 × 160 pixels 2 , 192 × 192 pixels 2 , and 256 × 256 pixels 2 ). We also extracted non-overlapping patches from unstitched B-scans. The B-scans patches had dimensions of size 156 × 118 µm 2 , 313 × 235 µm 2 , 352 × 264 µm 2 , 469 × 353 µm 2 , 625 × 470 µm 2 , and 781 × 588 µm 2 (32 × 32 pixels 2 , 64 × 64 pixels 2 , 72 × 72 pixels 2 , 96 × 96 pixels 2 , 128 × 128 pixels 2 , and 160 × 160 pixels 2 ). Fibroids were hand-labeled in 3D Slicer [ 58 ] (version 5.8.1) according to pathological annotations. Texture, statistical, and frequency-based features were derived from both en face and B-scan patches. The texture features included gray-level co-occurrence matrix (GLCM) based features contrast, correlation, energy, and homogeneity. Additionally, the variance, skewness, kurtosis, and entropy were extracted. Finally, the frequency content within 4 concentric circles within the patches was quantified using a Gaussian filter [ 59 ]; these features are Fourier P1, Fourier P2, Fourier P3, and Fourier P4. Fourier P1 is the smallest circle, and Fourier P4 is the largest. Optical features were extracted exclusively from the B-scan patches. The mean and SD of the attenuation coefficient map were extracted from patches with lateral sizes of 156 µm, 313 µm, 352 µm, 469 µm, 625 µm, and 781 µm (32 pixels, 64 pixels, 72 pixels, 96 pixels, 128 pixels, and 160 pixels) by performing iterative linear fitting [ 60 ] with consideration of confocal gating [ 61 ]. The attenuation coefficients were extracted from the detected surface until the A-line mean intensity reached a specified threshold. All features were extracted using MATLAB 2024B. Statistical differences between distributions were evaluated using nonparametric analysis of variance (ANOVA) and receiver operating characteristic (ROC) curve analysis; results were considered significant if p < 0.05. Due to our small sample size, we tested whether the medians of the grouped features were statistically different. More specifically, we used the Kruskal-Wallis test to compare ranks and Dunn’s test post hoc for pairwise comparisons. In the ROC curve analysis, the positive samples are the medians of the fibroid features, and the negative samples are the medians of the normal features. All analyses were performed using Prism (version 11.0.0, GraphPad Software, LLC). Patches were fully one-class, i.e., no patches with multiple labels were extracted.