EMFET

Research use only. EMFET classifies individual, pre-cropped blood-cell images (from the NIH malaria single-cell dataset or similarly structured data) as Parasitized or Uninfected using a handcrafted 5-feature pipeline and a classical ML classifier. It does not detect cells in a whole blood smear, estimate parasitemia, identify parasite species/stage, or produce a patient-level diagnosis. It has not been validated as a clinical or diagnostic device.

Install

pip install emfet

Quick start: run the app

emfet

Launches the bundled Streamlit demonstration app in your browser — upload a pre-cropped cell image, see the step-by-step feature-extraction pipeline, and get a prediction from the bundled, pre-trained model. No separate model file or dataset needed to try it.

Quick start: use the library

from emfet.features import extract_features

features = extract_features("sample_cell.png")
# array([n_spots, max_spot_area, total_spot_area, spot_saturation, gray_std])

What's in the package

• emfet.app / the emfet console command — a Streamlit demonstration app, bundled with a pre-trained model, launchable with zero setup.
• emfet.features — the feature-extraction pipeline (Gray World color normalization, adaptive spot detection, morphological cleaning, HSV saturation), driven by a FeatureConfig so every processing stage is a configuration, not a code fork.
• emfet.data — dataset loading and patient-ID resolution for patient-grouped evaluation (required to avoid patient-level leakage on datasets with multiple images per patient).
• emfet.evaluate — a patient-grouped nested cross-validation harness with a built-in leakage guard, a full diagnostic-metrics suite, a permutation significance test, and paired Wilcoxon comparisons for ablation studies.
• emfet.ablation — stage/feature ablations and classical baseline descriptors (color histogram, LBP, GLCM, HOG) for comparison under the identical evaluation protocol.
• emfet.benchmark — controlled latency/memory/CPU measurement, reporting model size, load memory, and peak inference memory as distinct, explicitly-labeled quantities.

Labels

Parasitized = 0, Uninfected = 1 (emfet.data.LABEL_NAMES). The positive class for every metric in emfet.evaluate is Parasitized (emfet.data.POSITIVE_CLASS).

Security note

Loading a pickled/joblib model executes arbitrary code during deserialization. Only load model files from sources you trust.

Full project, training/evaluation scripts, and reproducibility harness

This package is the importable library. The full research repository — CLI scripts for patient-grouped nested cross-validation, ablation sweeps, benchmarking, a Streamlit demonstration app, tests, and the manuscript this software accompanies — is at https://github.com/abkafi1234/cell_image.

License

MIT.