Abstract
Microstructure segmentation is essential for quantitative materials analysis; however, supervised deep learning demands substantial annotation, whereas general-purpose foundation models such as the Segment Anything Model (SAM) offer limited domain-specific semantic control. This paper presents MatStudio, a human-in-the-loop framework for microstructure segmentation that is proposed and implemented end to end in this work. MatStudio couples an interactive workflow for batchwise micrograph annotation and model adaptation with a dual-head convolutional architecture and SAM-guided boundary refinement. The loop combines sparse supervision with SAM-assisted labeling, task-specific training, and iterative batch-level correction, typically converging within two to three cycles.. The network comprises a shared encoder initialized from a pretrained backbone and two decoders: a UNet-style segmentation head that jointly predicts class labels and pixelwise uncertainty, and a prototype branch that measures texture similarity against a memory bank of learnable prototypes. An uncertainty-aware training objective couples label estimation with confidence; uncertainty weighting emphasizes prototype learning in ambiguous regions, and purity filtering restricts prototype updates to phase-interior features to mitigate boundary contamination. At inference, decoder outputs are fused in a spatially adaptive manner according to local confidence. We further introduce SAMFusion, which treats SAM as a morphology-aware proposal source and accepts proposals that are consistent with the fused prediction under a hierarchical small-to-large activation criterion, yielding improved boundary fidelity without case-specific tuning of SAM prompts. Experimental evaluation on Ni-alloy micrographs and public benchmarks shows consistent gains relative to interactive scribble-based segmentation and SAM-centric reference methods.