Structured Conditioning and Multiscale Generation of Cell Microscopy Images for Segmentation Data Augmentation
Year: 2026
Type: article
Abstract: Generating cell microscopy images across sources requires balancing appearance quality, conditional response, and structural preservation. We developed a multiscale conditional diffusion model (MSCDM) for text-conditioned generation without spatial references and foreground generation given contours and backgrounds. Three text segments condition three diffusion scales. For foreground generation, parameter interpolation between a base model and a class-distribution teacher forms a shared backbone; after freezing it, teacher predictions supervise sparse feature modulation gated by reference and target classes. Evaluation covered five imaging configurations from four public datasets. Fréchet Inception distances were 53.729 for 10,000 text-generated images and 25.076 for 4000 same-class foreground resyntheses, the lowest in their respective comparisons, with reference-fidelity trade-offs. Evaluator-defined classification accuracy for fixed-anchor cross-class requests was 85.0% (153/180). Across two HeLa sequences, mixed real and same-class synthetic training data improved instance-matching mean average precision (mAP) in both transfer directions. The bidirectional macro-average gain over real-data training was 1.359 percentage points across three paired seeds, with matched updates and effective batch sizes. These findings support foreground appearance augmentation within existing real backgrounds, layouts, and annotations.
Source: Electronics
Authors: Bin Shen(Investigation, Methodology, Software, Writing – original draft), Hang Xu(Investigation, Methodology, Software, Writing – original draft), Zhen Gu(Funding acquisition, Methodology, Resources, Validation), Jiale Zhou(Software, Validation), Bingyong Yan(Supervision, Writing – review & editing), Huifeng Wang(Funding acquisition, Resources, Writing – review & editing)
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $2,896
Award: G7288879442
OpenAlex ID: W7236300521
