Researchers have developed a hierarchical ensemble inference pipeline to improve the accuracy of automated white blood cell classification, particularly in the presence of domain shifts. This method utilizes a memory-augmented approach with a DinoBloom backbone fine-tuned via LoRA and incorporates k-nearest neighbors retrieval at multiple stages. Tested on the WBCBench dataset for the ISBI 2026 challenge, the pipeline achieved a top-ten ranking based on macro F1-score, demonstrating its robustness in identifying critical rare cell subtypes like blast cells. AI
Summary written by gemini-2.5-flash-lite from 1 source. How we write summaries →
IMPACT Improves robustness of medical image classification models against real-world data variations.
RANK_REASON Academic paper detailing a new method for a specific classification task.