The world today is witnessing the AI revolution like no other. It has reached everyone's phone, with large language models (LLMs) that generate text and more recently, images and videos, becoming the main carrier of this large scale mass adoption of AI. However, alongside the LLM or generative AI, there is another quieter revolution happening. This has not so much to do with generating new text, images or videos, but more to do with the analysis of existing images and videos. Object detection, classification, segmentation, automated annotation and analyses is not something that one may need in everyday life. But, it has become absolutely revolutionary in scientific research and healthcare.
Extensive resources and immense efforts have also gone into the development of several deep learning algorithms for object detection and segmentation of biological image datasets. Deep learning applications in biological imaging rely on a mix of classical detection and segmentation models, each suited to different challenges posed by microscopy data. Two-stage detectors like the Faster R-CNN provide high accuracy for small objects such as cells and nuclei, while one-stage models like YOLO (You Only Look Once) and SAM (Segment Anything Model) enable faster, real-time analysis in applications ranging from bacterial cells (1-2 micron in size) to larger organisms like C. elegans (1-2 mm in size). Other segmentation models such as U-Net and Mask R-CNN allow pixel-level identification of structures even in dense or overlapping samples.
Therefore, today, there is a growing repertoire of AI models and they are becoming increasingly powerful, accessible and specialized. The possibilities are vast, and limited only by our imagination. The question is no longer whether AI can transform how we analyze scientific images and videos, but how far we are willing to take it. Scientists who embrace these tools today will not simply be adopting a new technology; they will be shaping the future of how scientific research is conducted. And, those who ignore it may find themselves trying to catch up with a revolution that has already begun.
Here are some latest models to help get started:
- YOLO: https://docs.ultralytics.com/models/yolo26
- SAM: https://docs.ultralytics.com/models/sam-3
- Faster R-CNN: https://docs.pytorch.org/vision/main/models/faster_rcnn.html
- Mask R-CNN: https://docs.pytorch.org/vision/stable/models/mask_rcnn.html
- U-Net: https://www.ultralytics.com/blog/a-guide-on-u-net-architecture-and-its-applications
- CellPose (for cells): https://www.cellpose.org/
- Stardist (for cells): https://stardist.net/
- DeepCell (for cells and nucleoids): https://www.deepcell.org/
- ScCamAge (for multimodal cellular dataset): https://github.com/the-ahuja-lab/scCamAge
- From our lab: Deep-Worm-Tracker (for C. elegans): https://github.com/knaticat/Deep-Worm-Tracker
- From our lab: Deep-Pose-Tracker (for C. elegans, from our lab): https://github.com/cebpLab/Deep-Pose-Tracker
Comments
Post a Comment