Abstract
Segmentation of medical images is crucial for accu-rate diagnosis, treatment planning and intervention. However, the scarcity of data and the variability of different imaging modalities, including MRI, CT, PET and ultrasound, pose a major challenge for deep learning models. This review discusses traditional and advanced data augmentation techniques used to improve segmentation performance. In addition to geometric transformations such as rotation and scaling, advanced approaches such as Generative Adversarial Networks (GANs) and superpixel-based augmentations are also evaluated. The review also highlights novel methods such as Local-and-Contour Aware Grid Mixing (LCAMix) and Hard and Soft Mixing (HSMix) that preserve anatomical boundaries and effectively augment datasets. The analysis concludes with practical insights for clinical applications and suggests integrating these techniques into data pipelines to increase the generalizability of models. Future research directions focus on DICOM-specific augmentation strategies and optimization of computational requirements to improve clinical applicability.
Nancy Kozah
PhD Student
My research focuses on designing deep learning-based data augmentation techniques for medical image segmentation.
Fadi Dornaika
Ikerbasque Research Professor
Ikerbasque Research Professor with expertise in computer vision, machine learning, and pattern recognition.