Accurate spermatozoa tracking is crucial for understanding fertilization and reproductive health and developing novel sperm-based diagnostics and therapies. This paper explores the application of YOLOv8, a state-of-the-art object detection model, for automated and robust spermatozoa tracking in microscopic videos. Our approach employs transfer learning, fine-tuning the pre-trained YOLOv8 model on the VISEM-Tracking dataset of labeled spermatozoa images. We evaluate the performance of the proposed method on a specific part of the dataset, which does not participate in the training step. By providing a reliable and eficient tool for automated spermatozoa tracking, this work paves the way for further research, advancements, and applications in the field of reproductive medicine.
Spermatozoa tracking in microscopic videos demands both laser-sharp detection and seamless connection across frames. This challenge divides computer vision algorithms into two main camps: the methodical two-stage approach and the lightning-fast one-stage approach.
The two-stage approach in object detection involves initially proposing candidate regions in
an image and subsequently classifying and refining those regions. This approach operates
with meticulous precision. Algorithms under this paradigm meticulously examine each frame,
carefully identifying individual sperm. Examples include advanced methods like DeepSORT
[
In contrast, the one-stage approach in spermatozoa tracking employs a single, high-speed model
to both detect and trace sperm in microscopic videos. YOLO (You Only Look Once), a renowned
object detection system, exemplifies this approach with its real-time eficiency [
By examining the VISEM-tracking dataset, and from the oficial implementation of the dataset
[
As discussed, regarding the compatibility of the dataset format with the YOLO model, we employed the YOLOv8 small (s) configuration, we conducted the training duration to 100 epochs to further perform the detection and tracking task.
Our decision to use YOLOv8 small and extend the training duration was driven by the need to strike a balance between sensitivity and adaptability. By setting the Confidence Threshold at 0.25 and the NMS IoU Threshold at 0.7, we aimed to achieve several objectives: • Balanced Sensitivity and Adaptability to Challenging Conditions: We intend our model to be sensitive enough to detect spermatozoa accurately, even in challenging conditions. With a lower Confidence Threshold, we want to ensure a higher recall rate and minimize the false negatives, as well as make the model more robust and reliable, which is crucial in medical applications. • Accuracy and Precision: At the same time, we aim to maintain a balance between accuracy and precision. This ensures that the detected spermatozoa are indeed spermatozoa and not false alarms. • Redundancy Removal: The NMS IoU Threshold at 0.7 allowed us to eliminate redundant or overlapping bounding boxes. This step is crucial for enhancing tracking accuracy, as it ensures that each spermatozoon is represented by a single bounding box, preventing multiple detections of the same sperm.
In summary, our approach with YOLOv8 small and the specified threshold values was designed to strike a careful balance between sensitivity, precision, adaptability, and tracking accuracy for the challenging task of spermatozoa detection and tracking in microscopy images, ultimately contributing to more accurate and eficient sperm analysis in clinical practice.
While YOLOv8s works well in accurate sperm identification and tracking, its inherent processing demands may not be readily compatible with the eficiency of system requirements of Task 2. Recognizing this potential limitation, we opted to prioritize Task 1’s accuracy and robustness, ensuring a dependable baseline for further improvement.
The table below displays our YOLOv8-s model submission using the IoU threshold set at 0.7.
Submission YOLOv8s+Conf@.25
Precision
The validation set shows a precision of 0.5 and a recall of 0.628, indicating our approach detects 62.8% of spermatozoa in microscopy images. However, the 50% false positive rate suggests room for model optimization. For insights into accuracy and robustness across overlap thresholds, we considered mAP50 and mAP50-95. mAP50 at 0.506 signifies a mean average precision of 50%, showcasing reasonable detection accuracy. mAP50-95 at 0.191 emphasizes the need for improved robustness, especially at stricter overlap thresholds.
In the evaluation of the sperm tracking task using YOLOv8, the performance metrics reveal
noteworthy insights. The HOTA (Higher Order Tracking Accuracy) [
Besides, the localization accuracy (LocA) of the model is also high at 70.7, indicating that it can accurately estimate the positions of the sperm cells. Moreover, our model performs inference with an FPS (frame per second) metric of 54.38 and an FLOPS (floating-point operations per second) score of 8,570,207,776, which indicates that our model is very competitive in terms of inference time.
YOLO provided some techniques for model augmentation like HSV augmentation, Image Flip, Image FixUp, Segment CopyPaste. These augmentation methods could potentially enhance the model’s performance. Moreover, utilizing tools like Ray Tune, used for parameter optimization could also refine the model’s performance. In the extension of our method, we can consider using these augmentation and hyperparameter tuning methods to improve the result returned by YOLOv8.
Even though the dataset is organized to be compatible with YOLO models, diferent methods can still be eficiently implemented to perform the sperm object detection and tracking task, resulting in the understanding of fertilization. In future work, we can find some experiments to perform transparent tracking and detection of spermatozoa with diferent methods, such as Deep CNN or Faster R-CNN to evaluate and compare the results between models.
Moreover, the result from this task can also be inferred to be used for the prediction of motility, the ability of an organism to move independently. And from that, we can observe whether a spermatozoon can "move forward", or can only move around in circles.
This research is funded by Viet Nam National University Ho Chi Minh City (VNU-HCM) under grant number DS2020-42-01.