SHMC-Net: A Mask-guided Feature Fusion Network for Sperm Head Morphology Classification
CoRR(2024)
摘要
Male infertility accounts for about one-third of global infertility cases.
Manual assessment of sperm abnormalities through head morphology analysis
encounters issues of observer variability and diagnostic discrepancies among
experts. Its alternative, Computer-Assisted Semen Analysis (CASA), suffers from
low-quality sperm images, small datasets, and noisy class labels. We propose a
new approach for sperm head morphology classification, called SHMC-Net, which
uses segmentation masks of sperm heads to guide the morphology classification
of sperm images. SHMC-Net generates reliable segmentation masks using image
priors, refines object boundaries with an efficient graph-based method, and
trains an image network with sperm head crops and a mask network with the
corresponding masks. In the intermediate stages of the networks, image and mask
features are fused with a fusion scheme to better learn morphological features.
To handle noisy class labels and regularize training on small datasets,
SHMC-Net applies Soft Mixup to combine mixup augmentation and a loss function.
We achieve state-of-the-art results on SCIAN and HuSHeM datasets, outperforming
methods that use additional pre-training or costly ensembling techniques.
更多查看译文
AI 理解论文
溯源树
样例
生成溯源树,研究论文发展脉络
Chat Paper
正在生成论文摘要