ThinkLIF: multi-scale auxiliary representation guided nuclear segmentation in IHC histopathology

Reliable analysis of tissue microenvironments is often constrained by the limited molecular information available in standard immunohistochemistry (IHC) and the high cost of multiplex immunofluorescence (mpIF). Although deep learning has advanced computational pathology, existing methods still struggle with staining heterogeneity and ambiguous cellular boundaries because they often lack explicit morphology-sensitive mechanisms for robust structural refinement. We propose ThinkLIF, an auxiliary-representation-guided framework for nuclear segmentation in Ki67-stained IHC images. The model generates stain-like auxiliary representations and integrates them with the original IHC input through a ConvNeXt-style segmentation head, GatedFusion, and a FiLM-based AuxSkipAdapter to improve multi-scale boundary refinement. Experiments on the BCData and DeepLIIF benchmarks show that ThinkLIF achieves highly competitive segmentation performance. On BCData, it obtains the best Pixel Accuracy, Dice, IoU, and AJI among the compared methods. On DeepLIIF, it achieves the best Dice, IoU, and IHC Quantification Difference, while also maintaining competitive AJI performance. Overall, the results support improved boundary delineation, instance separation, and cell-level Ki67 scoring consistency under heterogeneous histopathological conditions. ThinkLIF provides an effective way to improve nuclear segmentation from widely accessible IHC images by combining original morphology with auxiliary stain-like representations. Further validation is required before these auxiliary outputs can be interpreted as quantitative molecular surrogates.