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

J. T. Hagbe; Michel Emel

cs.CVSep 26, 2026

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

Authors: J. T. Hagbe, Michel Emel

Organizations: African School of Economics, Abomey-Calavi, Benin · African Institute for Mathematical Sciences (AIMS), Cameroon

Abstract

Automated malaria diagnosis from thin blood-smear microscopy could meaningfully reduce the burden on under-resourced laboratories, but a model that maximises accuracy on clean laboratory images fails badly the moment an inexpensive smartphone camera introduces sensor noise. This paper introduces MalariaCascade, a two-stage inference system in which a lightweight MobileNetV2 sentinel (2,225,153 parameters) makes confident classifications at low compute cost and escalates uncertain cases to an EfficientNet-B3 expert (10,697,769 parameters) that sees only clean, standardised images regardless of how corrupted the incoming frame is. Structural isolation of the expert stage, not learned robustness, is the mechanism. The sentinel is trained under a safety-score objective that places an explicit floor on Recall(Parasitised) (>=0.95) and Recall(Uninfected) (>=0.40) before precision is optimised, ensuring the checkpoint satisfies clinical safety constraints by construction. On the NIH Malaria Cell Images Dataset (27,558 cells), the cascade reaches Accuracy=0.9736, Recall(Parasitised)=0.9570, Precision(Parasitised)=0.9912, F1=0.9738, and AUROC=0.9955 on the clean test set. Under Gaussian sensor noise at full severity, cascade Recall(Parasitised) degrades by only 2.4 pp (0.9570 to 0.9329), while the flat single-model baseline collapses by 63.0 pp (0.9584 to 0.3281). McNemar's test confirms the cascade improvement is statistically significant (chi^2=11.14, p=0.00085). An ablation isolating the structural property shows that routing clean images to the expert is responsible for a 16.8 pp Recall(Parasitised) advantage under sensor noise relative to a cascade where the expert also sees corrupted inputs.

Figures & tables

Model	Accuracy	Rec_P	Prec_P	F1	AUROC	$\rho$
Baseline CNN	$0.9620$	$0.9570$	$0.9684$	$0.9627$	$0.9891$	—
Augmented CNN	$0.9594$	$0.9546$	$0.9656$	$0.9601$	$0.9870$	—
Flat MobileNetV2	$0.9637$	$0.9541$	$0.9744$	$0.9642$	$0.9934$	—
Sentinel-only ( $\tau=0.50$ )	$0.9715$	$0.9522$	$0.9916$	$0.9715$	$0.9955$	0.0%
Expert-only	$0.9782$	$0.9801$	$0.9773$	$0.9787$	$0.9967$	100.0%
MalariaCascade ( $\tau^{*}=0.45$ )	$0.9736$	$0.9570$	$0.9912$	$0.9738$	$0.9955$	49.7%

Table 1: Test-set performance on clean images for all evaluated models.

\rho

is the fraction of test samples routed to ExpertNet. “—” denotes models (Baseline CNN, Augmented CNN, Flat MobileNetV2) that are not cascade stages and for which

\rho

is not defined. Sentinel-only invokes ExpertNet 0% of the time; Expert-only invokes it 100% of the time. MalariaCascade uses

\tau^{*}=0.45

. AUROC uses sigmoid-probability scores.

Figure 1: SentinelNet training curves across three panels. Left : focal loss on the training (dashed) and validation (solid) sets, both decreasing monotonically. Centre : per-epoch Recall(Parasitised), Recall(Uninfected), and Precision(Parasitised) on the validation set; the dotted line marks the 0.95 safety floor. Right : Safety Score per epoch (equal to Prec_P when both recall constraints are met, else zero); the safety constraint is satisfied from the first epoch onward.

Figure 2: Five-model comparison on the clean test set (Expert-only omitted for visual clarity; its metrics appear in Table 1 ). Figure axis labels use “TL MobileNetV2” and “SentinelNet” for the models referred to as Flat MobileNetV2 and Sentinel-only respectively in the text. MalariaCascade achieves the highest Precision(P) among the multi-backbone configurations; Sentinel-only ( $\tau=0.5$ ) reaches marginally higher Precision(P) by applying a more conservative classification threshold.

Figure 3: Validation-set threshold sweep. Safety Score (Precision(P) when both recall constraints are satisfied, else zero) peaks at $\tau^{*}=0.45$ . The shaded region marks the 0.955 Recall(P) floor used for threshold selection.

Figure 4: Pareto frontier: Recall(Parasitised) versus expert load $\rho$ as the routing threshold $\tau$ varies. The selected $\tau^{*}=0.45$ (filled circle) lies on the frontier at the safety-constraint boundary.

$\tau$	Rec_P	Prec_P	F1	$\rho$	SS
0.35	$0.9693$	$0.9870$	$0.9781$	51.7%	0.9870
0.40	$0.9645$	$0.9889$	$0.9765$	50.5%	0.9889
0.45	0.9570	0.9912	0.9738	49.7%	0.9912
0.50	$0.9494$	$0.9941$	$0.9712$	49.1%	—
0.55	$0.9418$	$0.9945$	$0.9675$	48.6%	—

Table 2: Test-set threshold sweep (selected range). Safety Score (SS) = Prec_P when test Rec_P

\geq 0.95

, else ‘—’.

\tau^{*}=0.45

(bold) was selected on the validation set (val Rec_P = 0.9638, transfer gap = 0.0068).

Model	Corruption	Rec_P	Prec_P	F1	Accuracy
MalariaCascade	Clean	$0.9570$	$0.9912$	$0.9738$	$0.9736$
	Field	$0.9546$	$0.9917$	$0.9728$	$0.9727$
	Sensor	$0.9329$	$0.9945$	$0.9627$	$0.9630$
	Staining	$0.9532$	$0.9916$	$0.9720$	$0.9719$
Flat MobileNetV2 †	Clean	$0.9584$	$0.9726$	$0.9655$	$0.9649$
	Field	$0.9806$	$0.7907$	$0.8755$	$0.8573$

Table 3: Robustness evaluation at full corruption severity (

s=1.0

). Values for MalariaCascade reflect corruption applied to sentinel inputs only; ExpertNet receives clean images in all cascade conditions. † Flat MobileNetV2 denotes the same model as the “Flat MobileNetV2” row in Table 1 ; clean-row values here are from the robustness evaluation pipeline and may differ from Table 1 by

\leq\!0.005

due to an independent evaluation pass over the same fixed test set.

Figure 5: Grouped bar chart of Rec_P, Prec_P, F1, and Accuracy for MalariaCascade (red) and Flat MobileNetV2 (blue) across all four evaluation conditions at full corruption severity ( $s=1.0$ ). The sensor-noise group is the discriminating scenario: cascade Rec_P falls only marginally while the flat model collapses to 0.33.

Figure 6: Recall(Parasitised) versus corruption severity for MalariaCascade (solid red) and flat MobileNetV2 (dashed blue). Shading marks the cascade advantage ( $\mathrm{Cascade}-\mathrm{Flat}>0$ ). Sensor corruption (centre) shows the largest divergence.

Corruption	Rec_P	Prec_P	F1	eff- $\rho$
Clean	$0.9513$	$0.9796$	$0.9652$	49.7%
Field	$0.9764$	$0.8149$	$0.8884$	61.3%
Sensor	$0.2572$	$0.9963$	$0.4089$	13.2%
Staining	$0.9206$	$0.9828$	$0.9507$	47.9%

Table 4: H1 test: Flat MobileNetV2 evaluated at

\tau_{\mathrm{H1}}=0.570

, the highest threshold satisfying Rec_P

\geq

0.95 on the validation set (val Rec_P = 0.9500). eff-

\rho

is the fraction of test samples predicted Parasitised at that threshold. For reference, MalariaCascade at

\tau^{*}=0.45

achieves Rec_P = 0.9333 under sensor noise.

Condition	Clean	Field	Sensor	Staining
Normal cascade	0.9570	0.9546	0.9333	0.9532
Both-corrupted	0.9570	0.9428	0.7641	0.9461
Expert-only corrupted	0.9773	0.9589	0.7636	0.9650
Interaction	$-$ 0.0203	$-$ 0.0161	$+$ 0.0005	$-$ 0.0189

Table 5: Both-Corrupted Stress Test: Recall(Parasitised) under three evaluation conditions at full corruption severity (

s=1.0

). “Normal” applies corruption to sentinel inputs only. “Both-corrupted” applies the same corruption to both stages simultaneously. “Expert-only corrupted” applies EfficientNet-B3 to all 4,135 test samples under corruption, without routing. Interaction = Both-corrupted

-

Expert-only-corrupted: values near zero indicate additive failure; negative values indicate that routing degradation compounds Expert corruption.

Oracle comparison
	Rec_P	Prec_P	F1	$\rho$
Actual cascade	$0.9570$	$0.9912$	$0.9738$	49.7%
Oracle routing	$0.9773$	$0.9876$	$0.9824$	51.1%
$\Delta$	$+$ 0.0203	$-$ 0.0036	$+$ 0.0086	$+$ 1.4%

Table 6: Oracle routing and end-to-end routing bias at

\tau^{*}=0.45

. Top : oracle vs. actual cascade on clean test images; Oracle routes all true Parasitised cells to ExpertNet and applies SentinelNet at

\tau^{*}

to all true Uninfected.

\Delta

is Oracle

-

Cascade. Bottom : routing decomposition by true class; rates in parentheses are relative to the true-class count.

Variant	Accuracy	Rec_P	F1	AUROC	$\rho$	Sensor Rec_P
Full cascade ( $\tau^{*}\!=\!0.45$ )	$0.9736$	$0.9570$	$0.9738$	$0.9955$	49.7%	$0.9329$
Sentinel-only ( $\tau\!=\!0.50$ )	$0.9715$	$0.9522$	$0.9715$	$0.9955$	0.0%	$0.9289$
Expert-only	$0.9782$	$0.9801$	$0.9787$	$0.9967$	100.0%	—
H3 (no isolation)	same clean metrics as full cascade					$0.7646$

Table 7: Ablation results. Clean-data metrics (Accuracy, Rec_P, F1, AUROC) match Table 1 ; Prec_P is omitted for space (see Table 1 for Prec_P values). Metrics are reproduced here for direct within-table comparison. The rightmost column reports Recall(Parasitised) under sensor noise at full severity (

s=1.0

), computed by an independent evaluation pass in the pre-computation script. Expert-only is not evaluated under end-to-end corruption (“—”) because it has no routing gate and its inputs are always clean by definition. The H3 row tests structural isolation versus model capacity: when the expert is forced to receive the corrupted input instead of the clean pipeline image, Sensor Rec_P falls from 0.9329 to 0.7646.

Model	$\mathrm{Rec}_{\mathrm{P}}$	$\mathrm{Prec}_{\mathrm{P}}$	F1	AUROC	$\rho$
MalariaCascade ( $\tau^{*}{=}0.45$ )	$0.8197$	$0.9499$	$0.8800$	—	75.3%
Flat MobileNetV2 ( $\tau{=}0.50$ )	$0.9551$	$0.6981$	$0.8066$	$0.9038$	—
Expert-only (EfficientNet-B3)	$0.8283$	$0.9486$	$0.8844$	$0.9605$	100%
Sentinel-only ( $\tau^{*}{=}0.45$ )	$0.9768$	$0.6489$	$0.7797$	$0.9182$	0%
NIH P. falciparum reference (Table 1 )
MalariaCascade	$0.9570$	$0.9912$	$0.9738$	$0.9955$	49.7%

Table 8: Cross-dataset generalisation to BBBC041 ( P. vivax , Broad Institute / Kaggle). Evaluation: 2,452 Parasitised crops (ring, trophozoite, schizont, gametocyte) plus 2,452 randomly sampled Uninfected crops (balanced, seed 42) from 1,328 whole-slide images. Models applied without retraining at NIH-optimised thresholds. AUROC is not defined for MalariaCascade (two-stage system, no single output probability).

\rho

: fraction of cells routed to ExpertNet. NIH reference rows are from an independent evaluation pass; values may differ from Table 1 by

\leq\!0.005

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
EMFE: A lightweight, explainable machine learning framework for malaria cell classification
Aug 25, 2026Md Abdullah Al Kafi, Walayat Hussain, Mousumi Karmakar +2Medical Image Classification Classification
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 +1
On-Device Multi-Species Malaria Detection with Uncertainty-Calibrated Slide-Level Aggregation
Aug 9, 2026Idaya Seidu, Ahmed Tahiru Issah, Charles B. Delahunt +1Medical Diagnosis Object Detection
Carnegie Mellon University Africa, Kigali Innovation City, Kigali, Rwanda · University of Washington, Seattle, Washington, USA

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

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

On-Device Multi-Species Malaria Detection with Uncertainty-Calibrated Slide-Level Aggregation