Towards automated metaphase cell detection using foundation models: A SAM and DINO-based approach

Karyotyping is a widely used laboratory technique to detect abnormalities in chromosomes. Karyotyping consists of many steps, including collecting samples, cell culturing and metaphase cell detection. Cytogeneticists examine all the images manually and annotate metaphase cells that is tedious and time consuming after collecting the samples and creating cell cultures. Metaphase cells can be detected by using computer-aided techniques. In this study, Segment Anything Model (SAM) and Self-Distillation with No Labels (DINO) foundation models were employed for metaphase detection task to minimize training time and costs. Potential metaphase regions were detected using SAM foundation segmentation model and these regions were given to classification models to identify metaphase cells. ResNet-50, a convolutional neural network algorithm, Vision Transformer (ViT) and Cross-Covariance Image Transformer (XCIT) based backbones were used in metaphase cell identification step. The results show that foundation models can give promising results on metaphase cell detection as much as supervised deep learning-based models. As a result, while supervised XCIT model with small architecture exhibit 0.9966 True Positive Ratio (TPR) and self-supervised ViT base accomplish 0.9961 TPR.