En undersökning av Vision Transformers för kvantitativ cellanalys

Quantifying immune cell activity, especially how natural killer (NK) cellskill tumor cells, is important for advancing cancer immunotherapy. Today,microscopy-based cell analysis of these interactions is still mostly manual,which makes it slow and time-consuming. Automated methods exist but oftenstruggle with accuracy. This thesis explores whether Vision Transformer(ViT)-based models can help improve automated segmentation and detectionof these cell killing events, compared to more traditional convolutional neuralnetworks (CNNs).Four deep learning architectures were evaluated: DeepLabV3+ (CNN-based segmentation), TransUNet (CNN-transformer hybrid), Mask2Former(transformer-based segmentation), and YOLOv12 (transformer-based objectdetection). All models were trained and evaluated using multi-channelmicroscopy images showing NK and tumor cells, and tested on both in-distribution (ID) and out-of-distribution (OOD) data. Yolov12 achieved thehighest accuracy on ID data and performed well at detecting individualcells, while Mask2Former demonstrated strong generalization and consistentpredictions on more challenging OOD samples, likely benefiting from itstransformer-based design. DeeplabV3+ produced detailed segmentationmasks but sometimes suffered from oversegmentation, and TransUNet showedless stable performance overall. These findings indicate that transformer-basedmodels such as Mask2Former and YOLOv12 showed strong performance incell segmentation and detection. Future work with larger and more diversedatasets, while also including temporal data, could help further evaluate thebenefits of using transformer-based models for microscopy image analysis ofcells.