An image segmentation method and medium for densely populated regions of living cells
By generating predictive auxiliary modality images through a feature extraction network and optimizing dynamic weights, the undersegmentation problem of single-modal input in the segmentation of dense live cell regions is solved, achieving high-precision and low-cost live cell image segmentation, avoiding model degradation and high annotation costs.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SAIL SPACE (SUZHOU) INTELLIGENT TECHNOLOGY CO LTD
- Filing Date
- 2026-03-23
- Publication Date
- 2026-05-26
AI Technical Summary
Existing deep learning models suffer from problems in segmenting densely populated regions of live cells, such as undersegmentation under single-modal input, high cost of fine-tuning large models that easily leads to a decline in generalization ability, difficulty in obtaining paired multimodal auxiliary information during the actual inference stage, and existing pseudomodal prediction and splicing schemes are prone to feature attention shift and performance degradation.
By acquiring a single-modal main image and pairs of real auxiliary modal images, a feature extraction network is used to generate predicted auxiliary modal images. Through dynamic weight optimization and loss function adjustment, combined with a pre-trained large segmentation model, image segmentation is performed, achieving high-quality auxiliary modality generation and feature fusion under single modality.
Without relying on real multimodal input, it improves the segmentation recall and accuracy of densely populated live cell regions, avoids the high cost of manual annotation and the risk of model degradation, and achieves high-precision, low-cost single-modal live cell instance segmentation.
Smart Images

Figure CN122090447A_ABST