Comparative Analysis of Attention – Based and Convolution YOLO Architecture for Small Object Detection in Medical Imaging

Koguri, Rahul (2025) Comparative Analysis of Attention – Based and Convolution YOLO Architecture for Small Object Detection in Medical Imaging. Masters thesis, Dublin, National College of Ireland.

Preview	PDF (Master of Science) Download (567kB) | Preview
Preview	PDF (Configuration Manual) Download (625kB) | Preview

Abstract

The detection of small objects is a long-standing issue in the medical domain, since pathological areas tend to be less than one percent of the image size. Though methods in the YOLO family have transformed from convolutional networks to attention-based networks, there has been no comparison regarding their efficiency in the detection of small medical abnormalities. This research presents the first comprehensive analysis comparing attention-based YOLO variants (YOLOv11m with C2PSA, YOLOv12m with Area Attention) against CNN-based models (YOLOv8m, YOLOv10m) in four different medical imaging techniques: brain tumor MRI, liver histopathology, chest radiography, and bone fracture detection.

The work assesses sixteen model and dataset pairs based on standardized training procedures and a set of evaluation metrics such as mean average precision, the trade-off between precision and recall, and inference speed. The results show that architectural supremacy is context-dependent and thus not generic across datasets and tasks. CNN architectures outperform in well-defined pathologies and when the number of training samples is low, and the attention-based architecture has an advantage in identifying diffuse boundaries typical of histopathological images. The efficiency-optimized architectures appear to be enough to handle large pathological features.

The results call into question the assumptions surrounding the superior function of the attention mechanism, instead demonstrating that the characteristics of the dataset, such as the size distribution and definition of the object boundaries, and the amount of available training data, are more informative of the choice than the assumed complexity of the model. The contribution offers evidence-based recommendations for practitioners regarding the choice of architectures in detection tasks for medical imaging applications and the significance of matching architectural capabilities to corresponding medical imaging needs.

Item Type:	Thesis (Masters)
Supervisors:	Name Email Qayum, Abdul UNSPECIFIED
Uncontrolled Keywords:	MRI; CNN; MRI; YOLO; Attention
Subjects:	Q Science > QA Mathematics > Electronic computers. Computer science T Technology > T Technology (General) > Information Technology > Electronic computers. Computer science R Medicine > Healthcare Industry Q Science > Q Science (General) > Self-organizing systems. Conscious automata > Machine learning
Divisions:	School of Computing > Master of Science in Data Analytics
Depositing User:	Ciara O'Brien
Date Deposited:	08 Sep 2026 08:43
Last Modified:	08 Sep 2026 08:43
URI:	https://norma.ncirl.ie/id/eprint/9872

Actions (login required)

View Item