Compact INT8 malaria-cell classification: distillation, low-false-positive discrimination, and CPU inference trade-offs
Small malaria-cell classifiers require evaluation of false-positive behavior as well as model size. We combined an initial four-configuration study and a bounded six-configuration follow-up on one patient-separated National Library of Medicine (NLM) split. Three-seed validation selected distilled CompactMP and UltraLightSqueezeNet. Full-integer 8-bit (INT8) models were compared at unchanged validation-selected thresholds, with patient-cluster uncertainty and matched processor profiling. On the reused NLM test, mean sensitivity/specificity were 95.94%/96.87% versus 94.58%/97.14%. The sensitivity difference was +1.362 percentage points (conditional 95% interval -0.259 to +3.319); lower specificity meant the original improvement criterion failed. Secondary normalized partial area under the receiver operating characteristic curve (pAUC) over false-positive rates 0–0.05 favored CompactMP by 0.04462. Frozen transfer to 85,486 annotated BBBC041 Plasmodium vivax crops yielded mean sensitivity/specificity of 93.60%/50.59% versus 86.61%/52.28%; the external INT8 normalized-pAUC difference was +0.28671 [+0.26648, +0.30758] (exploratory field-cluster interval). Absolute normalized pAUC fell from 0.93116/0.88654 internally to 0.36784/0.08113 externally; both frozen thresholds transferred poorly. CompactMP's INT8 file was 46,224 bytes versus 42,696, with MacBook invocation medians 0.312 versus 0.382 ms. Validation and resource benefits did not establish overall predictive or clinical superiority.
Authors
- S. M. Abdulla Al Mamun
Institutions
- University of Manchester (GB)
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-19
- DOI
- https://doi.org/10.5281/zenodo.24374543
- Primary Topic
- Digital Imaging for Blood Diseases
- Type
- preprint