A deep learning-based computer-aided determination method for dominant follicle identification and antral follicle...

To develop and evaluate a multi-stage computer-aided determination (CAD) method for automated antral follicle count (AFC) and dominant follicle identification using pelvic magnetic resonance imaging (MRI), aiming to overcome the limitations of MRI-based follicular analysis, including high annotation cost and dependence on manual interpretation.

A total of 417 T2-weighted MRI slices from 124 patients were selected by radiologists, ensuring ovarian visibility and excluding poor quality images. The proposed CAD method incorporates a You Only Look Once version 11 detection model, which accurately selects MRI slices containing ovarian structures. This is followed by LCR-UNet, a novel segmentation model that integrates a lightweight atrous spatial pyramid pooling module and a cascaded decoder (CR-decoder) with channel shuffle convolutional block and residual block.

The model achieved a Dice similarity coefficient of 0.8571 for follicle segmentation on MRI images. Based on pixel area ranking and ellipse fitting, dominant follicle identification reached an accuracy of 92.86% and an F1-score of 0.8929. Additionally, size-based filtering enabled automated counting of antral follicles with a precision of 82.91%.

The proposed CAD method approaches the performance of experienced radiologists while reducing manual workload, offering a reliable tool for ovarian reserve and fertility assessment.