Instance-based data augmentation for data insufficiency improvement | |
| Author | Nattawach Sataudom |
| Call Number | AIT Thesis no.CS-24-04 |
| Subject(s) | Computer vision Image processing Computer science |
| Note | A thesis submitted in partial fulfillment of the requirements for the degree of Master of Science in Computer Science |
| Publisher | Asian Institute of Technology |
| Abstract | Data insufficiency presents a significant challenge in medical image analysis, particularly for CNN-based models which require large, annotated datasets. This thesis introduces a novel data augmentation technique called Instance-Based Data Augmentation (IBDA) to address this issue by augmenting object instance areas. The IBDA method includes object removal using the LAMA-inpainting model and instance augmentation through horizontal and vertical flipping.The main objective of this research is to propose the IBDA technique. Additionally, this research experiments with the Segment Anything Model (SAM) to determine if it can auto mate pixel-level annotation in medical imaging. The study evaluates IBDA’s effectiveness in enhancing CNN-based model performance and compares IBDA with traditional augmen tation techniques. Experiments on blood smear images from the ALL-IDB dataset and a private dataset from Phramongkutklao Hospital focused on Acute Lymphoblastic Leukemia (ALL) pre-screening applications.Results indicate that SAM’s zero shot with centered instance selection post-processing method provides the most accurate pixel annotations. IBDA, especially when combined with standard augmentation, significantly improves the performance of YOLOv8-based models in instance segmentation and object detection tasks.Without any augmentation, the base line performance of instance segmentation tasks was 89.9% mAP50, while the combined augmentation strategy improved this to 91.2%. In object detection tasks, the baseline per formance was 85.0% mAP50, which increased to 93.3% with the combined augmentation strategy.This study offers a cost-effective solution for medical image data preparation, improving diagnostic model accuracy. Future research will focus on optimizing IBDA,exploring additional augmentation methods, and applying these findings to other domains facing data insufficiency. |
| Year | 2024 |
| Type | Thesis |
| School | School of Engineering and Technology |
| Department | Department of Information and Communications Technologies (DICT) |
| Academic Program/FoS | Computer Science (CS) |
| Chairperson(s) | Dailey, Matthew N.; |
| Examination Committee(s) | Mongkol Ekpanyapong;Chaklam Silpasuwanchai; |
| Scholarship Donor(s) | Government of Thailand; |
| Degree | Thesis (M. Sc.) - Asian Institute of Technology, 2024 |