Speaker
Description
Many existing scientific workflows for segmentation-based data acquisition rely on intensive manual labour of domain experts to label samples as out-of-box Machine Learning models do not generalise well. Different frameworks adopt different data formats, evaluation metrics, and configuration interfaces, resulting in slower experiment setups and need for intermediary steps for result comparison.
We introduce a Pipeline to simplify and accelerate the finetuning and evaluation process to produce custom benchmarks and models for 2D image datasets for instance segmentation. The Pipeline is a unified tool for standardised data handling, experiment configuration, finetuning, inferencing and evaluation, supporting three state-of-the-art segmentation frameworks – Detectron2, Ultralytics YOLO (v8/v11), and the Segment Anything Model (SAM). Each stage of the pipeline can be executed via command line interfaces, with seamless switching between models. Designed with modularity and extensibility, the Pipeline allows integration of new models and datasets without restructuring the entire pipeline. We validate the Pipeline with a case study using a multi-magnification level CryoEM dataset. Overall, this work promotes reproducible benchmarking and enables more transparent and efficient model comparison.
