SPERM DETECTION, TRACKING, AND CLASSIFICATION FOR BULL SPERM MOTILITY MEASUREMENT

• Main Author: Hidayatullah, Priyanto
• Format: Dissertations
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/53882
• Institution: Institut Teknologi Bandung

Demand for beef is increasing in Indonesia as its population grows. However, beef production is not keeping pace with the needs of the community. Importing beef to meet this deficit cost Rp. 4.27 trillion in 2016 and spending on imported meat increases year on year. Artificial insemination is the single most applicable substitute technology to increase cattle production. High-quality frozen semen is needed for artificial insemination. Previously, the demand for frozen semen for artificial insemination was also met by imports. The Lembang Artificial Insemination Center (BIB) was given a government mandate to provide quality frozen semen in a sufficient quantity for artificial insemination. To achieve this, BIB needs to improve the performance of its main duty, the inspection of semen before it is frozen. Semen undergoes both macroscopic and microscopic examinations. At the Lembang Artificial Insemination Center, about 40% of fresh semen samples are removed each day because they do not meet the quality criteria. Of the semen that is discarded, 99% fails the microscopic examination. Though microscopic examination can be done manually, this has several disadvantages: high levels of subjectivity, intravariability and intervariability; the time required; and the observer’s eye fatigue. According to the SNI 2017 semen inspection guidelines, one of the main purposes of microscopic examination is to measure motility, which is the focus of this study. Obtaining reliable, motile bull sperm requires three steps: fast and accurate detection, tracking, and classification. Digital image processing is commonly used to detect sperm. This method is slow, has limited detection results, and is vulnerable to artifacts. Deep learning-based object detection methods have become popular among researchers due to their reliability and speed. One state of the art deep learning-based method is YOLOv3. However, it has not been found useful for examining bull sperm. This method requires a huge quantity of annotated sample data to prevent overfitting and achieve good accuracy. YOLOv3 also has a fairly large architecture that reduces the speed of training and testing. This study proposes the use of DeepSperm500, a model with a smaller architecture than YOLOv3 and hyperparameters optimized for speed and accuracy. DeepSperm500 achieved accuracy of 94.01 mAP in the test dataset, 6.5 mAP points higher than YOLOv3, at a speed of 66.98 fps (3 times faster than YOLOv3). An obstacle in multi-sperm tracking is tracking error caused by detection failure. This study proposes the use of Tracking-Grid and average angle of sperm motion to predict the sperm that were not detected. This method has been proven more effective than methods that use sperm head angles because non-progressive-motile sperm movements often do not follow the sperm head. Tracking-Grid has also been found effective in tracking fast-moving sperm. Tracking-Grid and average angle of sperm motion reduce change of identity (ID switch) and increase accuracy to (MOTAL) 69.4 overall. This is 16.2 points higher than the state of the art Deep SORT. The speed achieved is 32.36 fps, 1.4 times faster than Deep SORT. In classification to determine which sperm cells are progressive-motile, the majority of researchers use a single Computer Assisted Sperm Analysis (CASA) parameter with a static threshold value. This is effective for motile-progressive sperm classification, but less reliable in identifying non-progressive-motile sperm such as vibrating or floating sperm. This study proposes conducting classification using a support vector machine (SVM) with three CASA parameters: VCL, VSL, LIN. This is the Bull Sperm Progressive Motility Classifier (BSPMCSVM3CASA). Experimental results show that BSPMCSVM3CASA’s accuracy is 92.28%; 0.71-3.89 points higher than other classification methods. This study contributes to the body of knowledge on using DeepSperm500 for sperm detection, average angle of sperm motion and Tracking-Grid for multi-sperm tracking, and BSPMCSVM3CASA for motility classification of bull sperm. The experimental results show the proposed methods achieve greater accuracy and speed than previously used methods.