Assessment of ovarian follicular reserve according to ultrasound data based on machine learning methods

BACKGROUND: Ovarian reserve reflects a woman's ability to successfully realize reproductive function. The assessment of ovarian reserve is an urgent task for clinical practice [1] and is important in scientific research. The use of computerized diagnostic image processing methods can accelerate and facilitate the performance of routine tasks in clinical practice. Their use in retrospective data analysis for scientific purposes allows to increase the objectivity of the study and supplement it with auxiliary information [2]. The issue of ovarian localization and follicle segmentation on ultrasound images has been previously investigated in other works. For instance, Z. Chen et al. [3] employed the U-net model to identify follicles on ultrasound images. Similarly, V.K. Singh et al. [4] addressed a related problem using a variant of U-net, namely UNet++ [5], which has gained considerable traction in the field of medical image analysis [6]. AIM: The study aimed to develop machine learning models for analyzing ovarian images obtained from an ultrasound machine. MATERIALS AND METHODS: An open dataset with a labeled ovary region was used for pre-training ovarian segmentation and follicle detection models. Subsequently, the dataset, which contains marked-up ovarian and follicle regions, was employed for training and testing. It encompasses a total of approximately 800 examples from 50 unique patients. The localization of follicles in an ultrasound image is a challenging task. To address this, the designed detector system was divided into two parts: ovary segmentation and follicle detection within the selected region. This approach allows the model to focus on a region where there are no other organs and various ultrasound artifacts that can be falsely perceived as the object under investigation. For the purpose of ovarian segmentation, the UNet++ architecture [5] was employed in conjunction with the ResNeSt encoder [8], which incorporates the SE-Net [9] and SK-Net [10] attention mechanisms. The object detection model is employed to identify the location of follicles within the ovary, as it enables precise enumeration of the number of follicles, even in the presence of overlapping structures, a capability that the segmentation model lacks. In our study, we used the YOLOv8 model [11]. Furthermore, data preprocessing has been employed to enhance the quality of model predictions. This has involved the identification and removal of regions with auxiliary information, the reduction of noise, and the augmentation of data. RESULTS: Two ovarian localization models are presented based on the results of this study. The first model is a segmentation model with an IoU quality of at least 50%. The second model is a detection model with a mAP quality of at least 65%. A third model is a model for follicle detection with subsequent follicle counting. This model has an MAPE error not exceeding 35%. CONCLUSIONS: The study resulted in the proposal of a method for applying machine learning techniques to the task of analyzing ultrasound images. The developed segmentation and detection models reduce the time and errors in analyzing ovaries and follicles in the images. The use of an attention mechanism and data preprocessing improves the quality of the models. The neural network for follicle detection provides follicle counting, even when follicles overlap.