EMFE: A lightweight, explainable machine learning framework for malaria cell classification

Md Abdullah Al Kafi; Walayat Hussain; Mousumi Karmakar; Sumit Kumar Banshal; Ahmed Al Marouf

cs.CVAug 25, 2026

EMFE: A lightweight, explainable machine learning framework for malaria cell classification

Authors: Md Abdullah Al Kafi, Walayat Hussain, Mousumi Karmakar, Sumit Kumar Banshal, Ahmed Al Marouf

Organizations: Multidisciplinary Action Research (MARS) Lab, Daffodil International University, Dhaka, Bangladesh · Artificial Intelligence for Decision Excellence (AIDX) Lab, Peter Faber Business School, Australian Catholic University, North Sydney, Australia · Department of Computer Science and Engineering, Alliance University, Bangalore, India · Department of Medicine, Faculty of Medicine & Dentistry, University of Alberta, Edmonton, Canada

Abstract

Automated malaria diagnosis from stained blood-smear microscopy is dominated by deep convolutional neural networks that are accurate but computationally expensive, poorly interpretable, and rarely validated with patient-level rigor. We present EMFE (Efficient Mathematical Feature Extraction), a five-feature framework for classifying single red-blood-cell images as parasitized or uninfected using Gray World color normalization, adaptive green-channel thresholding, morphological spot detection, and classical machine learning. Using the NIH LHNCBC malaria dataset (27,558 images from 200 patients), we evaluate Random Forest, Histogram Gradient Boosting, and Support Vector Machine classifiers under patient-grouped nested cross-validation (K_outer=20, K_inner=3), ensuring that cells from each patient remain within a single fold. The optimized Random Forest achieves 94.6% pooled out-of-fold accuracy (95% CI [93.6, 95.7]), corroborated by an untouched 40-patient holdout test (94.3%) and a patient-level permutation test (p<0.001, 1,000 permutations). Ablation experiments quantify the contribution of individual features and pipeline stages. Hardware-matched comparisons with retrained DenseNet121, ResNet50, and MobileNetV2 models assess the accuracy-efficiency trade-off. Synthetic perturbations characterize three failure modes, while explainability analysis identifies spot saturation as the dominant discriminative feature. Patient-level aggregation further quantifies sensitivity-specificity trade-offs and false-positive accumulation. These results demonstrate a statistically rigorous, interpretable, and computationally lightweight alternative to deep learning, while explicitly quantifying its limitations.

Figures & tables

Fig 1: Representative sample images from the dataset. Parasitized cells (containing visible Plasmodium chromatin, ring forms, or trophozoites) and uninfected erythrocytes, illustrating the range of staining intensity, illumination, and background artifact present in the raw data.

Fig 2: Comparative visualization of the feature-extraction process. Original image, color normalization, isolated green channel, raw adaptive threshold, morphologically cleaned spot mask, and final detected contours, for an uninfected cell (top) and a parasitized cell (bottom).

Classifier	Search space
SVM	$C\in[10^{-1},10^{2}]$ (log-uniform); kernel $\in\{$ RBF, linear $\}$ ; $\gamma\in\{$ scale, auto $\}$
Random Forest	$n_{\text{estimators}}\in[50,300]$ ; $\text{max\_depth}\in[3,20]$ ; $\text{min\_samples\_split}\in[2,10]$
Hist. Gradient Boosting	learning rate $\in[0.01,0.2]$ (log-uniform); $\text{max\_depth}\in[3,15]$ ; $\text{max\_iter}\in[50,400]$

Table 1: Bayesian hyperparameter search spaces.

Parameter	Value
Dataset	NIH LHNCBC, 27,558 images, 200 patients (150 infected / 50 control)
Class balance	13,779 Parasitized / 13,779 Uninfected (exact)
Excluded images	0
Outer folds ( $K_{\text{outer}}$ )	20, StratifiedGroupKFold , patient-grouped
Inner folds ( $K_{\text{inner}}$ )	3, StratifiedGroupKFold , patient-grouped
Hyperparameter search	Bayesian optimization, $n_{\text{iter}}=15$ per outer fold

Table 2: Experimental setup.

Metric	Random Forest	Hist. Gradient Boosting	SVM
Accuracy	0.9463 [0.9360, 0.9572]	0.9462 [0.9362, 0.9568]	0.9463 [0.9358, 0.9579]
Balanced accuracy	0.9463 [0.9358, 0.9564]	0.9462 [0.9356, 0.9561]	0.9463 [0.9354, 0.9569]
Sensitivity	0.9335 [0.9169, 0.9495]	0.9316 [0.9147, 0.9472]	0.9320 [0.9149, 0.9484]
Specificity	0.9590 [0.9498, 0.9682]	0.9607 [0.9521, 0.9694]	0.9607 [0.9509, 0.9704]
PPV	0.9579 [0.9277, 0.9674]	0.9596 [0.9296, 0.9685]	0.9595 [0.9301, 0.9690]
NPV	0.9352 [0.9167, 0.9572]	0.9336 [0.9153, 0.9552]	0.9339 [0.9149, 0.9566]

Table 3: Pooled out-of-fold metrics under patient-grouped nested cross-validation (

n=27{,}558

; 95% CI across the 20 outer folds).

Fig 3: Pooled out-of-fold confusion matrix. Random Forest predictions across all 20 patient-grouped outer folds ( $n=27{,}558$ ), with counts and row-wise percentages for each true class.

Fig 4: ROC and precision-recall curves. Pooled out-of-fold ROC curves (left) and precision-recall curves (right) for all three classifiers, computed from the 27,558 pooled out-of-fold predictions of the patient-grouped nested cross-validation. AUC and average precision (AP) values match Table 3 ’s ROC-AUC and PR-AUC columns exactly, as both are computed from the identical prediction set.

Metric	Random Forest	Hist. Gradient Boosting	SVM
Accuracy	0.9429 [0.9263, 0.9582]	0.9424 [0.9255, 0.9581]	0.9440 [0.9277, 0.9591]

Table 4: Held-out patient-level test set (160 train / 40 test patients; 21,915 / 5,643 cells; 95% patient-cluster-bootstrap CI).

Descriptor	Accuracy	F1	$\boldsymbol{\Delta}$ Accuracy
Color histogram [ 23 ]	0.9311	0.9299	$-0.0149$
Local Binary Patterns [ 16 ]	0.7061	0.7078	$-0.2284$
Gray-Level Co-occurrence Matrix [ 15 ]	0.6448	0.6516	$-0.2917$
Histogram of Oriented Gradients [ 17 ]	0.8567	0.8591	$-0.0832$

Table 5: Baseline-descriptor comparison (Random Forest, reduced 5-fold protocol) against the full EMFE pipeline (accuracy 0.9460, F1 0.9453).

Stage ablations
Variant	Accuracy	F1	Wilcoxon $\boldsymbol{p}$
No color normalization	0.9435	0.9431	0.4375
Otsu threshold (vs. adaptive)	0.9381	0.9351	0.1875
Fixed global threshold	0.9430	0.9425	1.0000
No morphological opening	0.9430	0.9421	0.0625
Leave-one-feature-out

Table 6: Stage-ablation and leave-one-feature-out results (Random Forest, reduced 5-fold protocol, default configuration: accuracy 0.9463, F1 0.9456).

Perturbation	Accuracy	F1	Sensitivity	Specificity
Brightness, dimmer ( $\times 0.7$ )	0.9415	0.9409	0.9068	0.9782
Brightness, brighter ( $\times 1.3$ )	0.9240	0.9248	0.9106	0.9381
Contrast, reduced ( $\times 0.7$ )	0.5150	0.6791	1.0000	0.0036
Contrast, increased ( $\times 1.3$ )	0.9279	0.9286	0.9137	0.9428
Hue shift, $10^{\circ}$	0.9421	0.9425	0.9254	0.9596
Hue shift, $20^{\circ}$	0.9454	0.9459	0.9302	0.9614

Table 7: Full synthetic-perturbation robustness sweep.

Model	Accuracy	F1	Size	Latency (CPU)	Latency (GPU)
EMFE (Random Forest)	0.9463	0.9456	6.3 MB	4.6 ms	N/A (CPU-only)
DenseNet121	0.9698	0.9695	27.1 MB	197.4 ms	35.7 ms
ResNet50	0.9676	0.9674	90.0 MB	105.4 ms	11.7 ms
MobileNetV2	0.9670	0.9666	8.7 MB	17.3 ms	4.8 ms

Table 8: Deep-learning baselines , retrained under the identical patient-grouped protocol as EMFE’s headline evaluation (pooled over 20 outer folds;

n=27{,}558

Comparison	Metric	Median $\boldsymbol{\Delta}$	Folds won by DL model	Wilcoxon $\boldsymbol{p}$
DenseNet121 vs. EMFE	Accuracy	$+0.0198$	20 / 20	$1.9\times 10^{-6}$
DenseNet121 vs. EMFE	F1	$+0.0189$	20 / 20	$1.9\times 10^{-6}$
ResNet50 vs. EMFE	Accuracy	$+0.0171$	20 / 20	$1.9\times 10^{-6}$
ResNet50 vs. EMFE	F1	$+0.0178$	20 / 20	$1.9\times 10^{-6}$
MobileNetV2 vs. EMFE	Accuracy	$+0.0177$	20 / 20	$1.9\times 10^{-6}$
MobileNetV2 vs. EMFE	F1	$+0.0169$	20 / 20	$1.9\times 10^{-6}$

Table 9: Paired Wilcoxon signed-rank test : each deep baseline vs. EMFE, matched per fold (20 folds, identical partition).

Fig 5: Per-feature distributions by class. Distribution of each of the five EMFE features, split by true class, on the held-out test set.

Feature	Importance
Spot saturation ( $S_{\text{max}}$ )	$0.2670\pm 0.0056$
Total spot area ( $A_{\text{total}}$ )	$0.0341\pm 0.0025$
Max spot area ( $A_{\text{max}}$ )	$0.0301\pm 0.0016$
Number of spots ( $N_{\text{spots}}$ )	$0.0098\pm 0.0010$
Gray standard deviation ( $\sigma_{\text{gray}}$ )	$0.0021\pm 0.0009$

Table 10: Permutation feature importance (30 repeats, mean

\pm

std, held-out test set).

Fig 6: Partial dependence. Partial dependence of the predicted Parasitized probability on each of the five features individually, with the other four held at their observed distribution.

Fig 7: Representative prediction-category gallery. True-positive, false-negative, false-positive, and true-negative example cells from the held-out test set, with their corresponding feature-extraction pipeline visualization.

Feature	FN mean	TP mean	$\boldsymbol{p}$ (FN v. TP)	FP mean	TN mean	$\boldsymbol{p}$ (FP v. TN)
$N_{\text{spots}}$	5.57	8.16	$1.8\times 10^{-15}$	6.75	3.35	$9.4\times 10^{-16}$
$A_{\text{max}}$	72.21	170.87	$2.4\times 10^{-81}$	148.74	42.53	$9.9\times 10^{-44}$
$A_{\text{total}}$	171.03	367.28	$1.4\times 10^{-52}$	263.85	70.62	$6.4\times 10^{-38}$
$S_{\text{max}}$	19.06	104.76	$1.2\times 10^{-128}$	77.31	8.48	$2.1\times 10^{-53}$
$\sigma_{\text{gray}}$	19.73	20.46	$1.5\times 10^{-5}$	20.85	17.77	$2.0\times 10^{-17}$

Table 11: Failure-mode feature comparison (Mann-Whitney

U

, held-out test set).

Variant	Accuracy	F1	Wilcoxon $\boldsymbol{p}$
Block size 11 (default 21)	0.9520	0.9513	0.0625
Block size 31 (default 21)	0.9312	0.9303	0.0625
Block size 41 (default 21)	0.9212	0.9204	0.0625
Adaptive $C=2$ (default 5)	0.8916	0.8897	0.0625
Adaptive $C=8$ (default 5)	0.9514	0.9508	0.0625
Adaptive $C=12$ (default 5)	0.9542	0.9536	0.0625

Table 12: Threshold-parameter sensitivity (Random Forest, reduced 5-fold protocol) against the default configuration (accuracy 0.9463, F1 0.9456).

$\boldsymbol{\tau}$	Infected flagged	Sensitivity (95% CI)	Control flagged	Specificity (95% CI)
1	150/150	1.000 (0.975–1.000)	43/50	0.140 (0.070–0.262)
2	149/150	0.993 (0.963–0.999)	31/50	0.380 (0.259–0.518)
3	148/150	0.987 (0.953–0.996)	18/50	0.640 (0.501–0.759)
5	144/150	0.960 (0.915–0.982)	7/50	0.860 (0.738–0.930)
10	131/150	0.873 (0.811–0.917)	0/50	1.000 (0.929–1.000)
20	95/150	0.633 (0.554–0.706)	0/50	1.000 (0.929–1.000)

Table 13: Patient-level sensitivity/specificity under a minimum predicted-positive-cell threshold

\tau

(Random Forest).

True parasitemia range	$\boldsymbol{n}$ patients	Mean parasitemia	Sensitivity (95% CI)
0.000–0.158	38	0.086	0.947 (0.827–0.985)
0.158–0.328	37	0.225	1.000 (0.906–1.000)
0.328–0.583	37	0.440	1.000 (0.906–1.000)
0.583–0.902	38	0.761	1.000 (0.908–1.000)

Table 14: Sensitivity at

\boldsymbol{\tau=3}

, stratified by true parasitemia quartile (infected patients only).

Quantity	Value
Serialized model size	6.3 MB
Image decode latency (median)	0.47 ms
Feature extraction latency	0.61 ms
Prediction latency	3.16 ms
Full pipeline latency (mean)	4.6 ms
Resident memory increase (model load)	28.1 MB

Table 15: Computational efficiency (500 timed repetitions, 50 discarded warm-up calls, single-threaded).

Explore similar work

MalariAI: A Label-Resilient Decoupled Framework for Universal Cell Segmentation and Explainable Stage Classification in Dense Malaria Blood Smears
Jul 1, 2026Kaysarul Anas Apurba, Md Hasibul Hasan, Mohammed Ali +1Explainable Medical Image Analysis Microscopy Image Analysis
Department of Electrical and Computer Engineering, North South University, Dhaka 1229, Bangladesh · Department of Computer Science and Engineering, International University of Business Agriculture and Technology (IUBAT), Dhaka 1230, Bangladesh
Safety-Constrained Cascade Inference for Robust Malaria Cell Classification Under Field Corruptions
Sep 26, 2026J. T. Hagbe, Michel EmelMedical Image Classification Classification
African School of Economics, Abomey-Calavi, Benin · African Institute for Mathematical Sciences (AIMS), Cameroon
Beyond Accuracy: Evaluating Efficiency, Robustness and Explainability in Deep Learning for Malaria Diagnosis
May 29, 2026Olivier Kanamugire, Kerol DjoumessiEfficient Neural Network Inference Image Corruption Robustness
African Institute for Mathematical Sciences, Kigali, Rwanda · Hertie Institute for AI in Brain Health, University of Tübingen, Germany

Jul 1, 2026, eess.IVJ/K move · Enter open · S save

MalariAI: A Label-Resilient Decoupled Framework for Universal Cell Segmentation and Explainable Stage Classification in Dense Malaria Blood Smears

Kaysarul Anas Apurba, Md Hasibul Hasan, Mohammed Ali, Tanzilur Rahman

Department of Electrical and Computer Engineering, North South University, Dhaka 1229, Bangladesh · Department of Computer Science and Engineering, International University of Business Agriculture and Technology (IUBAT), Dhaka 1230, Bangladesh

Explainable Medical Image Analysis Microscopy Image Analysis

Open paper AskPDF arXiv

Automated malaria diagnosis from blood smear microscopy is a critical challenge in global health AI; in resource-limited settings, the scarcity of expert microscopists remains the primary bottleneck to timely and accurate diagnosis. Three compounding failure modes prevent reliable clinical deployment of existing deep learning systems. First, end-to-end detectors treat unannotated cells as background during training, producing recall figures that are strongly influenced by annotation completeness rather than reflecting true cell recovery. Second, Non-Maximum Suppression tends to suppress valid detections in dense smear regions where infection counts matter most. Third, existing whole-slide detection pipelines lack per-cell spatial evidence for clinical audit, despite image-level explainability methods such as Grad-CAM having been applied to malaria image classification tasks. We present MalariAI, a two-stage decoupled framework that addresses all three failure modes in a unified pipeline. Stage 1 applies an annotation-agnostic distance-transform guided watershed algorithm to isolate every cell in a full 1600x1200 blood smear image, recovering 75.95% of ground-truth cells by centroid localisation across the 120-image NIH BBBC041 test set without any ground-truth input. Stage 2 fine-tunes EfficientNet-B0 with Focal Loss (gamma = 2.0, per-class inverse-frequency weights) on 64x64 crops, achieving 98.36% overall classification accuracy with 87.5% and 75.0% per-class accuracy on the rare schizont and gametocyte stages, compared to only 24.57% and 25.95% AP for a Faster R-CNN baseline on the same classes. Grad-CAM++ heatmaps generated per detected cell provide instance-level spatial evidence for clinical audit, enabling microscopists to verify model predictions at the individual parasite level without sacrificing classification performance.

Abstract

Explore similar work

MalariAI: A Label-Resilient Decoupled Framework for Universal Cell Segmentation and Explainable Stage Classification in Dense Malaria Blood Smears

Safety-Constrained Cascade Inference for Robust Malaria Cell Classification Under Field Corruptions

Beyond Accuracy: Evaluating Efficiency, Robustness and Explainability in Deep Learning for Malaria Diagnosis