AI-Assisted Cervical Cancer Screening Using VIA Imaging
This project was completed as a final course project and explored the application of computer vision methods to cervical cancer screening using Visual Inspection with Acetic Acid (VIA). VIA is a low-cost screening technique commonly used in low-resource settings, but its diagnostic accuracy depends heavily on visual interpretation and clinical experience.
The objective of the project was to investigate whether deep learning models could support more consistent interpretation of VIA images, with a focus on practical challenges such as limited labeled data, variable image quality, and clinically meaningful evaluation.
My Role & Contributions
I served as team lead and ML engineer, with responsibility for both technical direction and coordination.
Team leadership
- Organized and led regular team meetings to plan milestones and divide work
- Allocated tasks across preprocessing, segmentation, and classification components
- Reviewed all team members’ code to ensure consistency, correctness, and reproducibility
- Helped teammates debug model training, data handling, and evaluation issues
Model backbone pretraining
- Designed and implemented the backbone pretraining strategy for the project
- Focused on improving feature representations prior to downstream tasks
- Investigated transfer learning and self-supervised approaches to address limited labeled data
Pipeline support
- Assisted with the development of classification and segmentation pipelines
- Helped integrate pretrained backbones into downstream models
- Analyzed intermediate outputs and failure cases to guide architectural and training adjustments
Technical Focus
- Deep learning for medical image analysis
- Backbone pretraining and transfer learning
- Classification and segmentation under data-constrained settings
- Evaluation strategies aligned with screening-oriented objectives
Outcomes & Learnings
- Successfully coordinated a multi-component ML project from design to final evaluation
- Demonstrated the impact of backbone pretraining on downstream model performance
- Gained experience leading technical teams and reviewing production-style ML code
- Developed a deeper understanding of real-world constraints in medical imaging workflows
This project strengthened my interest in medical image analysis and reinforced the importance of thoughtful model initialization, collaboration, and clear technical communication in team-based ML work.