The transferability and targeted optimization of MedSAM on the universal cell segmentation challenge
This work delves into the crucial domain of cell segmentation within the AI4Science field, addressing the significant challenge of segmenting individual cells in tissue observations. Given the variability in cell morphology and limitations in imaging techniques, developing a universal, automated cell segmentation approach is paramount yet challenging. Our focus shifts towards optimizing models to manage errors effectively, ensure data transferability, and maintain consistency across different datasets. We center our investigation on MedSAM (Medical Segment-Anything Model), a specialized model for cancer cell segmentation, evaluating its performance against the benchmark model, Cellpose, across a diverse dataset from over 20 biological laboratories and more than 50 experiments. This comparative analysis aims to assess the consistency and adaptability of MedSAM, leading to two main pathways for further investigation: optimizing MedSAM through transfer learning and structural adjustments, and addressing unexpected errors through characteristic analysis and innovative optimization methods like diffusion models. The outcomes of this project are anticipated to offer profound insights into the comparative strengths and weaknesses of MedSAM versus Cellpose, contributing to the advancement of cell segmentation methodologies. By enhancing the accuracy, adaptability, and application breadth of segmentation models, this research aims to significantly impact biological research and the broader AI4Science domain, facilitating a deeper understanding of biological intricacies.
The project benefited from the support provided by Prof. Jesper N. Tegner, whose guidance was instrumental in its execution. Additionally, Dr. Sumeer Khan provided invaluable assistance with the research background and the methodology design.
The reorganized data and model data will be made available upon reasonable request.