Khon Kaen University researchers developed LeukSNN, a lightweight spiking neural network for detecting acute lymphoblastic leukemia in peripheral blood-smear images. It reported 99.91-100% accuracy across three public datasets while using 8% of the multiplication and 28% of the addition operations of the efficient state-of-the-art comparator.
Key findings
- Accuracy ranged from 99.91% to 100% across the three datasets, while multiplication and addition counts fell to 8% and 28% of the reference method. The claimed advantage is therefore computational as well as predictive, although operation counts are not equivalent to measured latency, memory use, or hardware energy.
Why this matters globally
If externally validated, such models could support screening where hematology expertise or computing resources are limited and may suit neuromorphic hardware. Any screening system must remain robust across cell types, staining protocols, devices, and populations beyond the training data.
Thai researcher contribution
The Khon Kaen University team developed the architecture and evaluation framework, demonstrating Thai research capability at the intersection of neuromorphic computing, image analysis, and oncology.
Limitations to consider
The evidence comes from public datasets, without reported prospective hospital testing, clinician comparison, or site-level external validation. Near-perfect accuracy may be sensitive to patient-level splitting, duplicate images, and preprocessing differences, which require scrutiny in the full paper.
Verify the original sources
Applied SciencesRead the original article↗DOI: 10.3390/app16136774