Empirical Evaluation of Integrating DINO Self-Supervised Learning with Segmentation Architectures for Cellular Images
Empirical Evaluation of Integrating DINO Self-Supervised Learning with Segmentation Architectures for Cellular Images
Hosna Khademfar1 Reza Yazdi2
1) East Golestan Institute of Higher Education,
2) Assistant Professor of Artificial Intelligence, Faculty of Engineering, Golestan University, Gorgan, Iran,
IOI Code
:
AAHD-ABKFFB
Publication :
National Conference on Information Technology, Nanotechnology, Artificial Intelligence and Technological Futures Studies(itnaf.ir)
Abstract :
Accurate cell segmentation in microscopic images is one of the key steps in biological and medical image analysis, but the reliance of deep learning models on labeled data limits their applicability in real-world settings. Self-supervised learning has been proposed as a strategy for leveraging unlabeled data; however, its effectiveness in biological imaging, particularly under data scarcity, remains not fully clear. In this study, we empirically evaluated the impact of integrating features extracted from the self-supervised DINO model with two feature extractors based on ViT and ResNet-50, alongside two segmentation architectures, U-Net and MaxSigNet. The experiments were conducted on the standard LIVECell dataset using 250 training images, 60 validation images, and 140 test images, and the performance of six different configurations was compared using IoU, Dice coefficient, and mean average precision (mAP). The results showed that the baseline U-Net achieved the best overall performance, with values of 0.79 for IoU, 0.90 for Dice, and 0.92 for mAP. Furthermore, adding DINO features, whether in the ViT-based or ResNet-50-based variant, did not yield a statistically significant improvement and, in some configurations, led to a slight drop in performance. Analysis of the results indicates that the domain gap between natural images and microscopic images, the mismatch between self-supervised learning objectives and the requirements of precise segmentation, and the simplicity of the feature integration strategy are among the main factors behind this behavior. By providing a transparent analysis of the negative results, this study offers a more realistic understanding of the limitations of directly transferring self-supervised models to cellular segmentation.
Keywords : cell segmentation self supervised learning dino u net maxsignet bioimaging livecell