(E-193) Machine-learned automatic detection of follicles in ovarian tissues visualized by optical coherence tomography

Introduction:: Rapid visualization and quantification of follicles in the ovarian tissue of young cancer patients could accelerate the fertility preservation therapy through effective ovarian tissue transplantation. Optical coherence tomography (OCT) is a non-invasive imaging method with a high spatial resolution to visualize follicles [1, 2]. However, no method has been established to accurately identify follicles from images with superimposed speckle noise inherent to OCT. In this study, we employed machine learning for the automatic detection of follicles from OCT images of ovarian tissue.

Materials and Methods:: Ovarian tissue images were acquired from 3 to 4 day-old mice using a full-field OCT system with a halogen lamp (Light-CT Scanner, LLTech Inc.). Images of 1024 × 1024 pixels were median filtered to reduce speckle noise and divided into multiple 64 × 64-pixel images. A convolutional neural network (CNN) was trained with a batch size of 32 and 30 epochs, utilizing 1024 images each, including and excluding the follicles, as depicted in Figure (left) [3]. The CNN classified whether each test image contained a follicle or not and the entire image of the ovarian tissue was again composed of tiles of test images. Comparing CNN’s detection to that of medical professionals allowed us to assess the detection rate and precision.

Results, Conclusions, and Discussions:: Figure (right) demonstrates an example output image. The yellow dots indicate the detections by the physician and red squares by the CNN and their detection rate and accuracy were 0.91 and 0.61, respectively. The visualized primordial follicles appeared as black shadows due to the absence of interfering signals but overlapping speckle noise made classification difficult. An appropriate machine learning platform and structure for ovarian image analysis, sufficient training data sets, and effective speckle noise removal will enhance detection rates and accuracy. Furthermore, OCT systems with near-infrared wavelength bands that can visualize deep tissue with high spatial resolution, such as the 1300 nm band, could be considered. Thus, we have non-invasively visualized murine follicles with OCT and automatically classified them with CNN. These findings enable three-dimensional image reconstruction of follicle distribution and may contribute to cryopreservation of follicle-rich ovarian tissues and tailor-made transplantation therapy.

Acknowledgements (Optional): : This research was partially supported by the Research Grant from the Keio Leading-edge Laboratory of Science & Technology.

References (Optional): : [1] Peters I.T. et al., Clin. Cancer Res., 2016, 22, 5506-5513.
[2] Takae S. et al., J. Assist. Reprod. Genet., 2018, 35, 627-636.
[3] Saito K. et al., Biomed. Phys. Eng. Express, 2020, 6, 065026.