Problem
Medical image classifiers often fail under domain shift due to differences in acquisition, staining, and imaging conditions.
What I Built
Reproduced MedViTV2 and trained on BloodMNIST + Czech WBC to study cross-domain behavior and robustness.
Architecture
- Local + Global feature perception stages
- KAN module to enhance non-linear representation
- Attention design intended to scale while preserving efficiency (DiNA)
Results
Accuracy (BloodMNIST)
98.86%
Accuracy (Czech WBC)
97.68%
F1 (Both)
~97–99%
Confusion matrices show strong diagonal dominance across both domains, with most errors concentrated between a small set of visually similar classes.
Confusion Matrix
Key Insights
DiNA in practice
The paper introduces DiNA to scale attention while controlling complexity. In my runs, the KAN-integrated transformer backbone performed strongly on these datasets; DiNA’s benefits may become more pronounced at larger scale and higher diversity.
Why KAN helped
KAN layers use spline-based activations that can model complex non-linearities effectively. On smaller medical datasets, this can contribute meaningful gains even when explicit local attention improvements are limited.