Automatic Follicle Counting From Ultrasound Images of Ovaries Using MARDSE-UNET Model.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Reproductive Endocrinology

Background:

• Ovarian structure detection in ultrasound is vital for gynecological and reproductive medicine.
• Automated systems can assist physicians in interpreting complex ultrasound images.
• Accurate follicle identification is key for diagnosing conditions like PCOS and infertility.

Purpose of the Study:

• To develop and evaluate a CNN-based object detector for segmenting and counting ovarian follicles.
• To enhance follicle detection performance using an integrated deep learning architecture.
• To provide a comprehensive tool for reproductive health diagnostics.

Main Methods:

• Development of the Multi-Attention Residual Dilated UNet with Squeeze and Excitation (MARDSE-UNet) model.
• Integration of residual UNet, dilated UNet, and squeeze-and-excitation blocks.
• Utilized the USOVA3D dataset with 5-fold cross-validation for rigorous testing.

Main Results:

• MARDSE-UNet achieved high performance: 98.69% accuracy, 97.89% precision, 97.7% recall, 86.97% F1-score, and 95.66% IoU.
• Novel preprocessing and post-processing stages improved noise reduction and feature extraction.
• The model outperformed traditional CNNs and state-of-the-art methods.

Conclusions:

• The MARDSE-UNet model demonstrates superior performance in automated ovarian follicle detection.
• The system offers a valuable tool for diagnosing various reproductive health conditions.
• Further development can enhance clinical applications in gynecological imaging.