No deep learning needed: handcrafted rules catch ~7 in 10 infected cells

MaziyarPanahi · x · 2026-11-02

Maziyar Panahi shares an honest post-mortem of a handcrafted-rules (not trained-model) experiment for detecting infected cells.

The first rule set classified only 104 of 160 samples into the right band; he then fixed the rules on 33 other patients and re-scored the 160.
Final result: the rules find about 7 in 10 of the infected cells the expert marked (70% sensitivity).
Data: U.S. National Library of Medicine blood smears, research use only.

He explicitly flags the "honest part"—candidly reporting the weak first version and the iteration process.

More from Research