A cell instance segmentation network for microscopy images with lossless feature preservation and dynamic perceptual calibration

The rapid advancement of cellular immunotherapy has generated an urgent demand for the high-throughput and high-precision quantitative analysis of T-lymphocyte functions, the realization of which hinges on the accurate segmentation of individual cell instances within complex microscopy images. However, existing methods struggle to cope with challenges such as minuscule target sizes, high morphological polymorphism, and severe cell overlapping induced by high-density cultivation. To address these issues, we propose HDC-Net, a cell instance segmentation network featuring lossless feature preservation and dynamic perceptual calibration, which systematically resolves the aforementioned difficulties through a collaborative mechanism. Specifically, the Space-to-Depth Convolution module replaces traditional downsampling with a lossless spatial-to-depth rearrangement to preserve the edge features of minuscule cells. Concurrently, the Large Selective Kernel Network backbone dynamically adjusts the receptive field to adapt to multi-scale morphological variations. Furthermore, the Efficient Multi-Scale Attention module performs cross-dimensional calibration during the feature fusion stage, thereby mitigating semantic confusion in overlapping regions. Experimental results demonstrate that HDC-Net achieves superior segmentation precision (a mAP50 of 0.928) and real-time inference speed (27.93 FPS) on our self-built T-lymphocyte microscopy image dataset. Ablation studies validate the effectiveness and collaborative gains of each individual module, while downstream task analyses further confirm its high accuracy and robustness in cell counting and confluence calculation. This study provides a reliable and efficient technological tool for the automated quantitative analysis of T-lymphocytes.