PaperlayerView exact revision on arXiv

Source-linked AI summary

MORPHA: Morphology-Constrained Training and the Limits of Cross-Acquisition Transfer in Low-Resource Malaria Microscopy

Favour Okechukwu Igwezeke, Chikodili Helen Ugwuishiwu, Joseph Uzochukwu Emesiani, Samuel Ifebuche Agada, Ekenechukwu Lilian Anozie, Mary Ofuru Kama, Adaobi Chiazor Emegoakor

arXiv:2609.05990v1cs.CV

TL;DR

Cross-acquisition transfer of morphology-based constraints remains unclear, although malaria models trained on one smear preparation encounter images from another. MORPHA uses measured stage statistics as a soft training constraint; it improves binary transfer and calibration, while the paper also maps settings where morphology does not transfer or generic regularisation performs better.

  • Problem

    Whether morphology-based supervision transfers across smear preparations is rarely tested, despite annotation regimes differing in richness and cost.

    Source: §1 · p. 2

  • Method

    MORPHA encodes stage-conditional statistics from BBBC041 as a soft training penalty across annotation regimes without changing architecture or inference.

    Source: §1 · p. 2

  • Results

    The constraint reduces the binary cross-dataset generalisation drop by 30.8% at negligible within-domain cost and improves calibration, while two standard regularisers exceed it on raw transfer.

    Source: §5.4 · p. 7, §6 · p. 8

  • Takeaways & Limitations

    MORPHA adds an interpretable consistency signal in units of deviation from measured morphology that the compared regularisers do not provide.

    Source: §6 · p. 8

    The evaluation uses single runs per condition, one backbone, and one thin-smear source per regime.

    Source: §6 · p. 9

Abstract

from arXiv · show

In low-resource malaria microscopy, a model trained on one smear preparation routinely meets images from another, and how well morphology-based constraints transfer across this acquisition gap is unclear. We study this on real African field microscopy from Uganda (Lacuna), asking where encoding measured parasite morphology as a training constraint improves cross-acquisition transfer and where generic regularisation suffices. We present MORPHA, a morphological consistency constraint that derives stage-conditional statistics from the stage-annotated BBBC041 dataset and penalises predictions that deviate from them. Defined uniformly across binary, object-level, and stage-aware regimes without changing architecture or inference, it shapes training in the binary regime. The detection regime is a mapped boundary. On transfer from thin-smear cells to thick-smear field images, the constraint reduces the binary-classification generalisation drop by 30.8% (F1 0.578 to 0.699) at negligible within-domain cost and lowers in-distribution calibration error by 49% (ECE 0.0162 to 0.0082). A content-free control applying the identical constraint to random statistics recovers less of the drop (25.3% vs 30.8%), indicating the measured content, not constraining alone, contributes to the gain. Two standard confidence regularisers exceed the constraint on raw transfer, locating where morphology adds value and where generic regularisation suffices. We map two deployment-relevant boundaries: thin-smear statistics do not transfer to thick-smear detection (trophozoite AP@0.50 falls to 0.000), and cross-acquisition pseudo-labelling fails before filtering applies. Together these yield a morphology-grounded consistency signal and evidence-based guidance for malaria dataset and model design in low-resource settings.

1 Introduction

Morphology-based supervision is rarely tested across smear preparations, despite differences in annotation richness and the cost of stage-aware labels. MORPHA uses measured stage statistics as a training constraint to study cross-acquisition transfer and its boundaries.

1 Introduction

  • Binary labels identify infection, while object-level and stage-aware boxes add localisation and developmental stage, with stage-aware annotation costlier to produce at scale.Developmental stage can carry prognostic information that binary labels cannot convey.Source: §1 · p. 2
  • MORPHA encodes BBBC041 stage statistics as a soft training penalty across three annotation regimes without changing architecture or inference.It shapes training directly in the binary regime and is a mapped boundary in detection.Source: §1 · p. 2
  • 30.8% reduction in the binary cross-dataset generalisation drop at negligible within-domain cost, with a 49% improvement in in-distribution calibration.Two standard regularisers exceed MORPHA on raw transfer.Source: §1 · p. 2
  • Thin-smear statistics do not transfer to thick-smear detection, and cross-acquisition pseudo-labelling fails before morphological filtering can apply.These boundaries indicate that thin-smear benchmarks do not proxy thick-smear field performance.Source: §1 · p. 3

2 Related Work

Prior malaria systems commonly treat morphology as an emergent feature, while most studies evaluate within one smear preparation. Stage-aware supervision and transfer across preparations remain less explored.

2 Related Work

  • Most malaria deep-learning systems treat morphology as an emergent feature; binary classifiers dominate, including Fuhad et al. and Boit & Patil.Fuhad et al. report 99.23% on NIH thin-smear data, and Boit & Patil report 97.68% on the same data.Source: §2 · p. 3
  • Stage-aware supervision is rarer and harder; Chaudhry et al. classify parasite type and life-cycle stage, while Abbas & Dijkstra report neighbour-stage confusions.MorphXAI injects morphological supervision into a detection decoder but evaluates within a single preparation.Source: §2 · p. 3
  • Nearly every study trains and tests within one smear preparation, leaving open whether morphological supervision transfers from thin to thick smears.Prior work also notes image-level split inflation and comparatively little thick-smear detection.Source: §2 · p. 3

3 Datasets

The study uses three public datasets with distinct annotation regimes and smear preparations, with Lacuna’s field images serving as the primary transfer target.

3 Datasets

  • Figure 1 contrasts NIH thin-smear cell crops, BBBC041 thin-smear field images, and Lacuna thick-smear African field microscopy.The caption identifies smear preparation as the source of visual domain shift.Source: Figure 1 · p. 3, §3 · p. 4
  • Lacuna provides 2,747 full-field thick- and thin-smear images from Uganda, captured by smartphone over a microscope eyepiece, and is predominantly thick-smear.Its field provenance makes it the primary transfer target.Source: §3 · p. 4
  • NIH provides 27,558 segmented thin-smear cells with balanced parasitized and uninfected binary labels.The dataset contains 13,779 cells in each class.Source: §3 · p. 4
  • BBBC041 provides stage-aware boxes and supplies the morphological statistics; the study uses 1,208 of its 1,328 images.The selected images are divided into 1,026 training and 182 validation images.Source: §3 · p. 4

4 Methods

MORPHA derives stage-specific feature distributions from BBBC041 and adds a soft Mahalanobis penalty to task loss, evaluating the constraint across binary and detection transfer regimes.

4.1 Morphological statistics

  • MORPHA derives stage-conditional Gaussian statistics from 79,672 BBBC041 instances across six classes using five morphology features.The features are box area, aspect ratio, mean and standard deviation of normalised intensity, and area relative to the mean uninfected-RBC box.Source: §4.1 · p. 5, Table 1 · p. 5

4.2 Morphological consistency loss

  • For each predicted positive region, the consistency score is its minimum squared Mahalanobis distance to a parasite-stage entry, using features available in that regime.For NIH single-cell images without boxes, the score uses image-level intensity and is weighted by parasitized probability.Source: §4.2 · p. 5
  • The total objective adds the morphology loss to binary cross-entropy or the YOLOv8 composite detection loss, with λ = 0.1 selected on NIH validation for transfer.For detection, the term uses aspect ratio and intensity over fixed ground-truth boxes, which supply no gradient.Source: §4.2 · p. 5

4.3 Architectures and uncertainty

  • The unchanged ResNet-50 classifier and YOLOv8-small detector are used with 20-pass Monte Carlo Dropout; calibration is measured by 15-bin ECE.Uncertainty is the variance of the parasitized probability, and the study also measures its Pearson correlation with the morphology score over predicted positives.Source: §4.3 · p. 6

4.4 Pseudo-labelling and training

  • BBBC041 detections label 9,645 NIH parasitized cells; stage-labelled predictions above 0.15 confidence and d(r) ≤ 0.237 fine-tune a multi-task ResNet-50 head.Binary models train for 50 epochs with AdamW, while detection models train for 100 epochs with SGD.Source: §4.4 · p. 6

5 Results

MORPHA reduces the binary cross-dataset generalisation drop while preserving within-domain performance, but its detection implementation maps a transfer boundary rather than testing the constraint. Calibration, control, pseudo-labelling, and weight-sensitivity results qualify where morphology helps and where other approaches or settings differ.

5.1 Within-dataset performance

  • On NIH, Binary-MC reaches F1 = 0.9729 against 0.9739 unconstrained; detection models are identical because the ground-truth-box term supplies no gradient.Stage-aware detection reaches mAP50 = 0.806, and within-Lacuna detection reaches 0.835; a prediction-level detection formulation is future work.Source: §5.1 · p. 6

5.2 Cross-dataset generalisation

  • 30.8% relative reduction in the generalisation drop: Binary-MC reaches Lacuna F1 = 0.6992 versus Binary-UC F1 = 0.5783, while within-domain F1 is 0.9729 versus 0.9739.The comparison is from thin-smear NIH cells to Lacuna field images; Table 2 defines drop as within minus cross-domain performance.Source: §5.2 · pp. 6–7, Table 2 · p. 7

5.3 Uncertainty and calibration

  • In-distribution ECE falls 49% (0.0162 to 0.0082), while out-of-distribution ECE falls 27% (0.4827 to 0.3517) but remains high.Both models remain overconfident on Lacuna; consistency quintiles show accuracy from 0.990 to 0.949, with a small d(r)–correctness correlation.Source: §5.3 · p. 7

5.4 Content and regulariser controls

  • Random statistics recover less of the generalisation drop than measured morphology (25.3% vs 30.8%), while label smoothing and a confidence penalty reduce it by 46.4% and 40.9%.The two regularisers exceed MORPHA on raw transfer and have lower OOD ECE; MORPHA instead provides an interpretable score of deviation from measured morphology.Source: §5.4 · p. 7, Table 3 · p. 8

5.5 Pseudo-labelling

  • Cross-acquisition pseudo-labelling yields only 17 candidate stage predictions from 9,645 NIH cells, five of which survive filtering, leaving the stage head effectively unsupervised.The detector’s domain mismatch leads nearly all NIH single-cell images to be labelled RBC; bridging the modes requires explicit domain adaptation.Source: §5.5 · p. 8

5.6 Sensitivity to λ

  • At λ = 0.01 transfer is worse than baseline; λ = 0.1 reduces the drop by 30.8% with negligible within-domain cost, while λ = 1.0 gives the lowest drop but lowers within-domain F1 by 0.010.The study selects λ = 0.1 for improved robustness without meaningful within-domain cost; λ = 1.0 suits settings prioritising robustness over benchmark accuracy.Source: §5.6 · p. 8

6 Discussion

MORPHA improves binary cross-dataset transfer and calibration, but its benefit is bounded by acquisition and task: thin-smear statistics do not transfer to thick-smear detection. The discussion also identifies experimental and modeling limits that constrain interpretation.

6 Discussion

  • The constraint reduces the binary transfer drop by 30.8% at negligible within-domain cost and improves calibration, while two standard regularisers exceed it on raw transfer.A content-free control indicates measured morphology, not constraining alone, contributes to the gain.Source: §6 · p. 8
  • Thin-smear statistics and models transfer partially in binary classification but not in detection, where a thin-smear detector essentially does not fire on thick smears.Thick smears are common in routine field screening.Source: §6 · pp. 8–9
  • Interpretation is limited by single runs per condition, constraint-only λ tuning, a gradient-free detection term, proxy features, weak morphology–uncertainty correlation, and one backbone and thin-smear source per regime.The authors describe results across single runs as directionally stable.Source: §6 · p. 9

7 Impact in Resource-Constrained Settings

Measured stage morphology can recover some robustness associated with richer annotation, while two zero-domain-knowledge regularisers achieve more. The authors identify CPU inference and field-acquired imagery as relevant to resource-constrained deployment and dataset design.

7 Impact in Resource-Constrained Settings

  • Measured stage morphology recovers part of the robustness richer annotation would provide, though less than two zero-domain-knowledge regularisers.Under budget limits, label smoothing or a confidence penalty is a sensible first step, with the constraint as a complementary interpretable addition.Source: §7 · p. 9
  • The classification model requires no GPU at inference and runs at 120 ms per image on a single CPU thread.The authors describe it as suited to microscopy workstations without dedicated accelerators.Source: §7 · p. 9
  • Across variants, models remained overconfident on Lacuna, making confident errors particularly risky where confirmatory review is unavailable.The authors state this risk applies regardless of the mitigation chosen.Source: §7 · p. 9
  • Thin-smear benchmarks cannot proxy thick-smear field performance, supporting deliberate inclusion of field-acquired imagery in future datasets.Source: §7 · p. 9

Related papers

Loading 2609.05990v1…