DC-UNet: Looking for follicles in the ovarian ultrasound images

Abstract

The auspicious initiation of human reproduction starts by releasing the ovum through ovulation within the ovary. Ceaseless monitoring of the female reproductive organs has now become essential for combating fertility-related issues and for successful assisted reproduction. Cases of infertility and demands for assisted reproduction in our modern liberated society are rapidly increasing. External or Transvaginal ultrasound imaging of the ovary provides us with vital information about the number, size, and position of the follicles in the ovary and their cumulative response to biological stimuli. Manual screening of thousands of USG images having lacs of follicles is an extremely strenuous job and prone to humane error. This paper propounded a new deep-learning architecture named Double Contraction-UNet (DC-UNet) which makes follicle segmentation fully automatic. This model restructured the U-Net architecture by introducing two contracting paths to segment the follicular object with higher accuracy. The model was trained and tested on approximately forty two thousand annotated ovarian 2D ultrasonography images extracted from USOVA3D Training Set 1. The proposed model outperforms the other U-Net-based state-of-the-art models when trained and tested on the same dataset. The proposed model has achieved an accuracy rate of 97.82%, a precision rate of 97.54%, a Recall value of 94.34%, an F1 Score of 95.91%, a Dice Score of 0.76, and a Jaccard Similarity Index of 0.59.

Keywords

• Ultrasonography
• U-net
• Double contraction layer U-net (DC-UNet)
• Follicle segmentation
• Dice score
• Jaccard similarity index