LI-SEMIS: High-precision instance segmentation for SEM images of lithium-ion batteries ternary materials

Accurate instance segmentation of electrode microstructures is essential for establishing processing-structure- property linkages and enabling automated quality assurance in lithium-ion batteries (LIBs) manufacturing. However, dense packing, strong adhesion, and multiscale morphology in LIBs ternary materials scanning electron microscopy (SEM) images often lead existing methods to oversegmentation or undersegmentation and blurred boundaries. We present LI-SEMIS (LIBs ternary materials SEM Instance Segmentation), a unified framework that augments the Segment Anything Model (SAM) with a UNETR style decoder for end to end, instance level prediction. The decoder jointly produces three complementary maps: foreground probability, distance to center, and distance to boundary, which are integrated via a watershed algorithm to yield pixel wise instance labels. Training adopts a joint optimization of Automatic Instance Segmentation (AIS) and Interactive Iterative Segmentation (IIS): simulated point/box prompts (with occasional reuse of previous masks) are progressively injected to co adapt the image/prompt encoders and the decoder. Inference integrates two key modules. Test Time Augmentation (TTA) fuses multi scale and flipped predictions to enhance global robustness, while the Corner Aware Fine Refinement (CAFRe) module performs high resolution re-prediction at image corners and employs a two dimensional Hann window feathering to smoothly blend local and global outputs, suppressing block artifacts and sharpening boundaries. On a dataset of 239 SEM images of LIBs ternary materials (split into 143/47/49 for train/val/test), LI-SEMIS consistently surpasses SAM and μSAM across ViT T/B/L backbones under SA50, SA75, and mSA. Ablations show CAFRe delivers substantial boundary fidelity and separation gains, TTA adds consistent robustness, and their combination is most effective; IIS further yields monotonic improvements with few user interactions exceeding the AIS baseline. LI-SEMIS provides a robust, generalizable solution for high precision microstructural analysis and scalable, automated inspection in LIBs ternary materials manufacturing.