Related Experiment Videos
The Simpler, the Better: A Lightweight Hierarchical Token Averager for Single-Label Microscopic Specimen Image
Objective:
Single-label microscopic specimen image classification (SMSIC) depends on subtle local morphology, yet many natural-image architectures devote substantial capacity to global interaction. We aim to develop an efficient local-first architecture tailored to this setting.
Methods:
We propose the Hierarchical Spatially Merged Token Averager (HSMTA), which reorganizes neighboring spatial elements and progressively compresses them into higher-level descriptors before a simple final average. HSMTA-SRR is introduced as an SRR-inspired, sparsity-regularized analytical counterpart for examining stage-wise activation selectivity and intra-class compactness.
Results:
Across SIPaKMeD, C-NMC, DSCC, and DeFungi, HSMTA achieves 98.1%, 76.6%, 94.5%, and 93.0% accuracy, respectively, with 1.1 M parameters and 0.4 GMACs. It outperforms five recent lightweight baselines on all four datasets and remains competitive with higher-capacity token-mixer models on DeFungi. Controlled comparisons with CNN+GAP, one-shot patch averaging, MIL-style pooling, and post-formation token mixers attribute the gain to hierarchical local descriptor formation rather than averaging alone. HSMTA-SRR further exposes stage-wise activation sparsification and intra-class compactness, providing model-centric diagnostics of representation formation.
Conclusion:
Local hierarchical descriptor formation followed by simple averaging provides a favorable accuracy-efficiency trade-off for morphology-dominant SMSIC tasks.
Significance:
HSMTA offers a compact specimen-level classifier for high-throughput digital microscopy workflows, where per-instance cost accumulates over many cells or local regions. The code is available at https://github.com/Dichao-Liu/HSMTA.
