ChromoNeXt: Morphology-Invariant 1-D Axial Modeling With Large-Kernel Convolutions for Chromosome Classification

Automated chromosome classification suffers from cross-domain generalization challenges caused by non-rigid distortions and imaging discrepancies. Existing two-dimensional vision models overfit morphological artifacts, while one-dimensional discrete sequence or time-series approaches disrupt the local spatial continuity of biological features. To address this issue, we reframe chromosome classification as a morphology-normalized one-dimensional spatial modeling task and propose ChromoNeXt, an ultra-lightweight convolutional architecture. First, we construct a standardized axial coordinate system via spatial resampling to achieve axial spatial normalization of banding profiles. The model then employs progressive large-kernel convolutions for spatial encoding, improving tolerance to local elastic misalignments and shifting the network’s focus from texture bias to a robust macro-shape bias along the canonical axis. Furthermore, an auxiliary masked reconstruction branch encourages contextual axial structure learning and enhances contextual reconstruction ability under severe blur. Extensive evaluations across four clinical datasets demonstrate that ChromoNeXt achieves strong in-domain accuracy and substantial improvements in source-only domain generalization. Crucially, in few-shot cross-modal adaptation scenarios, our model exhibits high data efficiency. ChromoNeXt thus provides a computationally efficient framework for chromosome classification in heterogeneous cytogenetic settings.