Semen analysis is crucial to determine men's fertility; however, microscope-based manual spermatozoa evaluation is time-consuming and costly. Therefore, it has become essential to develop computeraided-semen-analysis systems. To facilitate automated spermatozoa analysis, we propose a simple yet eficient framework for tracking sperms and predicting their motility. Diferent from existing methods, our proposed framework centralizes a new paradigm, dubbed sperm having a tail. We develop a novel tail-aware sperm detection model to advance the detection ability of dense, tiny, and transparent sperm cells. Furthermore, to enhance sperm tracking, a scene change detection technique is utilized to suppress identity assignment errors of similar sperms, resulting in improved sperm motility measurement. Experimental results show that our framework works well with an insignificant trade-of in execution time, which is suitable for the real-time clinical setting requirement.
Manual sperm quality assessment using microscopes is time-consuming and requires costly
experts with extensive training. In addition, the validity of manual sperm analysis becomes
unreliable due to limited reproducibility and high inter-personnel variations [
To promote automated sperm analysis, artificial intelligence has been developed and
substantially improved the assessment of the sperm [
The Medical Multimedia track of the MediaEval 2022 challenge, dubbed Medico 2022, tackles
the challenge of automatically tracking sperm cells in video recordings of spermatozoa [
Unlike previous works, which only focus on using diferent feature engineering techniques
and CNN architectures to improve the eficiency of CASA, we aim to tackle the challenges of
the small number of data and significant class imbalance. We propose a novel framework for
tracking sperms and predicting their motility. Our framework is concentrated on the
biologyinspired paradigm "sperm having a tail." Our contributions can be summarized as follows: (1)
We develop a simple yet eficient framework for tracking sperm cells and predicting sperm
motility. Our framework can meet the real-time clinical requirements. (2) We propose a novel
tail-aware sperm detection model. Our detection model can suppress uncertain detections near
the borders of video frames. We also empirically evaluate diferent implementation settings
of our detection model using YoloV7 [
Our preliminary detection results using YoloV5[
By exploiting the underlying properties of the data, we found that sperms on the border of
the videos tended to be ignored since annotators were uncertain (see Fig. 1(Right). Hence, we
propose suppressing the bounding box near the video frame borders. To achieve this goal, we
devise a biology-based solution that leverages the hypothesis that spermatozoa have tails (see
Fig. 1(Right). We propose a novel tail-aware sperm detection, which focuses on detecting
only sperm cells whose head is surrounded by a black line - the sperm tail. Particularly, we
combine various edge detection algorithms [
As in the dataset description [
We propose a simple yet eficient feature extraction followed by a shallow network to predict sperm motility (i.e., progressive and non-progressive). We first aggregate movement statistics of every tracked sperm resulting in features of each patient. For each sperm, we exploit three movement characteristics: (1) Total movement distance between neighboring frames; (2) Average velocity between two neighboring frames; (3) Distance between the first and the last location. These measurements are then averaged over the number of sperm, and the number of sperm cells is considered the fourth feature.
To avoid overfitting when training on limited data, we implemented a shallow network instead of using popular deep neural networks. The network contains one hidden layer with four hidden units, four input units, and three output units, followed by an extra soft-max layer. The network architecture corresponds to four feature inputs and three category outputs (e.g., motility rate, progressive rate, and non-progressive rate). We used L1 distance to calculate loss.
We tested our tail-aware sperm detection on the validation set. Table 2 demonstrates only results of the "sperm" category since the organizers concentrate on evaluating this weighty class. In our initial attempts, we naively suppressed all detections in the region near the borders, which led to a degradation in performance. Tail-aware sperm detection circumvents the mentioned problem by detecting sperms with tails, but it only improves precision. We argue that the model was not robust enough; thus, we added a confidence threshold to prevent it from discarding highly confident bounding boxes. Due to the class imbalance, we did not discard any cluster or small-pin-head sperm to prevent overuse of the tail-aware mechanism, which may lead to false positive results. This gives the class sperm a substantial 2.2% boost in terms of precision. Finally, we added the confidence threshold and the "only sperm filter" to our tail-aware sperm detection. The extra cost of tail-aware sperm detection is negligible since it only lowers the FPS by a margin of less than ten.
We verified our scene change detection (SCD) using the latest HOTA [
In this work, we proposed a novel tail-aware sperm analysis framework for transparent tracking of spermatozoa in the Medico 2022 challenge. Leveraging the biology-inspired paradigm "sperm having a tail," we proposed a tail-aware sperm detection to reduce false alarm detection of non-tail spermatozoa and a PSNR-based scene change detection to suppress incorrectly tracked trails. Experimental results on the Medico 2022 challenge demonstrated the potential of our proposed framework. We firmly believe that "sperm having a tail" is a strong cue for developing promising techniques for semen analysis.
In the future, we aim to investigate robust metrics for detecting scene changes to remove the performance dependent on Yolo. We also will study approaches that can overcome the limitation of currently implemented tail-aware sperm detection, such as hard-coded confidence thresholds and limited to regions near video frame borders.
Acknowledgement. This work is supported by Vingroup Innovation Foundation (VINIF) under project code VINIF.2019.DA19.