Automated Sperm Morphology Analysis Based on Instance-Aware Part Segmentation

TL;DR

This work tackles the challenge of automated, high-throughput sperm morphology analysis by addressing context loss and feature distortion in top-down instance segmentation and by introducing a robust centerline-based tail measurement method. It proposes an attention-based instance-aware segmentation network that refines preliminary masks with fused multi-scale features and edge information, plus an automated tail morphology measurement approach that reconstructs tail endpoints along the centerline gradient. Empirical results show substantial gains over state-of-the-art methods in instance segmentation and tail parameter accuracy, demonstrating the method's potential for reliable, quantitative sperm analysis in clinical settings. Together, these contributions enable faster, more accurate sperm morphology assessment with practical diagnostic impact.

Abstract

Traditional sperm morphology analysis is based on tedious manual annotation. Automated morphology analysis of a high number of sperm requires accurate segmentation of each sperm part and quantitative morphology evaluation. State-of-the-art instance-aware part segmentation networks follow a "detect-then-segment" paradigm. However, due to sperm's slim shape, their segmentation suffers from large context loss and feature distortion due to bounding box cropping and resizing during ROI Align. Moreover, morphology measurement of sperm tail is demanding because of the long and curved shape and its uneven width. This paper presents automated techniques to measure sperm morphology parameters automatically and quantitatively. A novel attention-based instance-aware part segmentation network is designed to reconstruct lost contexts outside bounding boxes and to fix distorted features, by refining preliminary segmented masks through merging features extracted by feature pyramid network. An automated centerline-based tail morphology measurement method is also proposed, in which an outlier filtering method and endpoint detection algorithm are designed to accurately reconstruct tail endpoints. Experimental results demonstrate that the proposed network outperformed the state-of-the-art top-down RP-R-CNN by 9.2% [AP]_vol^p, and the proposed automated tail morphology measurement method achieved high measurement accuracies of 95.34%,96.39%,91.2% for length, width and curvature, respectively.

Figures (6)

• Figure 1: (a) An image of stained sperm. Each sperm is to be segmented into five parts: acrosome, vacuole, nucleus, midpiece and tail. (b) Instance-aware part segmentation that not only distinguishes different sperm, but also segments parts for each sperm. (c) Ellipse and rectangle fitting for measuring sperm head and midpiece morphology parameters. (d) Centerline fitting for measuring sperm tail morphology parameters.
• Figure 2: (a) Context loss due to bounding box cropping. (b) Feature distortion due to resizing in ROI Align. (c) Endpoints mislocated in Steger-based methods. (d) The normal of endpoints is mislocated due to the influence of gradient from the intersecting edge.
• Figure 3: The structure of our proposed attention-based instance-aware part segmentation network. The convolutional backbone and FPN first extracts features from the input image then rescales extracted features to multi-scale. Next, the preliminary segmentation module generates instance-level parsing masks as in top-down methods. Finally, the refinement module refines preliminary generated masks by merging features extracted by FPN through the attention mechanism. Besides, the edge information is utilized to better separate boundary between adjacent sperm.
• Figure 4: (a) Normal lines intersect with opposite edges ($E_{11}, E_{12}$) and intersecting edges ($E_{21}, E_{22}$) for correctly detected point $P_1$ and mislocated point $P_2$. Correspondingly, the gradient $G_{11}, G_{12}$ for $E_{11}, E_{12}$ are close to parallel, while gradient $G_{21}, G_{22}$ for $E_{21}, E_{22}$ are not. (b) The gradient for center point $P$ can be decomposed to $G^\prime_x$ and $G^\prime_y$ along the line direction and perpendicular to the line direction. According to yang2008robust, $G^\prime_y=0$ (center point's gradient perpendicular to the line direction is zero); therefore, the center point's gradient direction is along the direction of line.
• Figure 5: Qualitative comparisons of instance-aware part segmentation networks. PGN (bottom-up method) has low instance distinction and cannot separate intersecting sperm parts. RP-R-CNN (top-down method) has distorted features and contexts outside a bounding box are cropped out. In comparison, our proposed network achieves the best prediction by refining preliminary segmented masks to retrieve lost contexts outside the bounding box and fix distorted features.