Cell Segmentation by Adaptive Decision Fusion in Time-Lapse Microscopy Image Sequences
Abstract
Cell segmentation in time-lapse microscopy images is crucial for cell behaviour analysis, and it is required for biological research and medical applications. However, it is a challenging task because of the low contrast, halo artifact, and frequent cell overlap. Manual cell segmentation is not only time-consuming but also requires expert knowledge and is prone to human error. Deep learning models have strong performance on cell segmentation, but their performance depends on the dataset and imaging conditions, which motivates the use of ensemble fusion strategies to improve robustness. This work proposes an Adaptive Decision Fusion (ADF) framework for cell segmentation to improve the accuracy and robustness. This framework works with probability maps produced by segmentation models and updates weights based on the model’s performance. The ADF performance is evaluated against Baseline, which is equal-weight averaging, and Majority Voting Fusion (MVF) across the six experimental configurations on three Cell Tracking Challenge (CTC) datasets. These experiments are designed to evaluate ADF based on various model quality and weight update frequency. The results show that ADF outperforms both Baseline and MVF in nearly all experiments across all three datasets. Online weight adaptation mechanism identifies and suppresses poorly performing models while amplifying the contribution of more reliable ones. This approach demonstrates robustness for cell segmentation across varying cell morphologies and imaging modalities, and has potential for future applications to advancements in medical research.