Kusumaningrum Ambarwati Prasetyo, Anindya (2025) Estimasi Abnormalitas Pada Spermatozoid Menggunakan Deep Learning Dengan SSD - MobileNet. Undergraduate thesis, UPN Veteran Jawa Timur.
![]() |
Text (Cover)
18081010098.-cover.pdf Download (900kB) |
![]() |
Text (Bab 1)
18081010098.-bab1.pdf Download (28kB) |
![]() |
Text (Bab 2)
18081010098.-bab2.pdf Restricted to Repository staff only until 6 August 2027. Download (894kB) |
![]() |
Text (Bab 3)
18081010098.-bab3.pdf Restricted to Repository staff only until 6 August 2027. Download (312kB) |
![]() |
Text (Bab 4)
18081010098.-bab4.pdf Restricted to Repository staff only until 6 August 2027. Download (770kB) |
![]() |
Text (Bab 5)
18081010098.-bab5.pdf Download (35kB) |
![]() |
Text (Daftar Pustaka)
18081010098.-daftarpustaka.pdf Download (90kB) |
Abstract
Spermatozoid morphology analysis is a crucial component in male fertility evaluation, with the condition of the acrosome being a primary indicator. The acrosome, a cap-like structure on the sperm head, releases enzymes essential for fertilization; thus, its damage or absence causes infertility. However, manual evaluation is subjective and time-consuming. This research aims to design and evaluate an automated object detection system using a deep learning method to identify and classify abnormalities in spermatozoa from microscope imagery, with a focus on acrosome morphology. The system was implemented using a Convolutional Neural Network (CNN) architecture, specifically SSD- MobileNetV2 FPNLite, through a transfer learning approach. The research methodology involved several testing scenarios to find the optimal model configuration. Tests were conducted by varying the number of training steps (5,000, 7,500, 10,000, and 20,000 steps) on the VISEM video dataset, as well as testing the impact of using a combined dataset (VISEM and MHSMA). Performance evaluation was conducted quantitatively using the mean Average Precision (mAP) metric and qualitatively through manual calculation of Precision, Recall, and F1-Score metrics. The results revealed that data annotation consistency is more critical than quantity. On the single dataset, a clear overfitting phenomenon was identified, where the model trained for 5,000 steps exhibited the best practical performance with the highest F1-Score of 61.29%, despite its mAP being only 5.3%. Further training steps significantly degraded the model's generalization ability. A major discrepancy was also found between automated and manual evaluations, highlighting the importance of qualitative validation for complex object detection cases. It is concluded that the model at 5,000 steps is the most effective and balanced for the practical application of detecting sperm morphology abnormalities. Keywords: Object Detection, Spermatozoid, Acrosome, Sperm Morphology, CNN, SSD-MobileNetV2, Transfer Learning.
Item Type: | Thesis (Undergraduate) | ||||||||||||
---|---|---|---|---|---|---|---|---|---|---|---|---|---|
Contributors: |
|
||||||||||||
Subjects: | T Technology > T Technology (General) | ||||||||||||
Divisions: | Faculty of Computer Science > Departemen of Informatics | ||||||||||||
Depositing User: | Anindya Kusumaningrum Ambarwati Prasetyo | ||||||||||||
Date Deposited: | 06 Aug 2025 07:05 | ||||||||||||
Last Modified: | 06 Aug 2025 07:05 | ||||||||||||
URI: | https://repository.upnjatim.ac.id/id/eprint/41606 |
Actions (login required)
![]() |
View Item |