PaperlayerView exact revision on arXiv

Source-linked AI summary

Towards Transparent Diagnostics: Investigating Architectural Trade-offs and Explainability in Malaria Detection

Suman Kunwar, Avishek Dangol

arXiv:2609.31682v2cs.CV

TL;DR

Microscopic malaria diagnosis is the gold standard but is labor-intensive and requires expertise. The study benchmarks CNN architectures using patient-grouped NIH blood-smear images and examines predictions with Grad-CAM, SHAP, and LIME. ResNet18+TTA achieves the highest reported accuracy, while pruning reduces its inference time and improves accuracy.

  • Problem

    Microscopy is the gold standard for blood-smear diagnosis but is labor-intensive, time-consuming, and requires experts.

    Source: §1 · p. 1

  • Method

    The study benchmarks malaria-classification architectures using patient-grouped evaluation and applies Grad-CAM, SHAP, and LIME to examine predictions.

    Source: §3.3 · p. 6, §5 · p. 11

  • Results

    97.96% accuracy and 0.9966 AUC were achieved by ResNet18+TTA; pruning reduced inference time from 14.55 ms to 10.65 ms and improved accuracy by an additional 0.21%.

    Source: §4 · pp. 7, 9

  • Takeaways & Limitations

    MobileNetV2 offers the smallest model size and fastest inference, while ResNet18+TTA achieves the best accuracy and AUC but has longer inference time.

    Source: §5 · p. 11

    Clinical or user validation was not performed, and real-world effectiveness remains to be tested.

    Source: §4 · p. 11

Abstract

from arXiv · show

More than 80 countries have reported malaria cases with 610 thousand deaths and are projected to increase. Identifying malaria early and accurately helps save lives and effective way to diagnose malaria is through microscopic methods that are labor intensive and require experts with special equipment. Deep learning (DL) has shown promising results in medical diagnosis. Here, we explored various DL models: ResNet18, MobileNetV2, EfficientNet-B2, VGG19 and proposed model ResNet18+TTA (ResNet18 backbone with modified classification head and test time augmentation) for detecting malaria presence using blood smears taken from the NIH Malaria dataset. Our experiment shows MobileNetV2 achieved 96.85 % accuracy with smallest model size (8.49 MB) and fastest inference (1.35 ms). The ResNet18+TTA model achieved 97.96 % accuracy, 0.996 AUC with longest inference time (13.32 ms). Larger architecture outputs a larger model size with moderate accuracy. Upon further pruning, ResNet18+TTA model gained a slight improvement in accuracy and reduced inference time. GRAD-CAM, SHAP and LIME provide explainable AI (XAI) insights into model predictions, using explanation agreement and divergence to evaluate predictive reliability.

1. Introduction

Malaria remains a major health burden, while microscopy—although the gold standard for blood-smear identification—requires time, expertise, and equipment. The study motivates automated diagnosis by identifying leakage and explainability gaps in prior deep-learning evaluations.

1. Introduction

  • In 2024, 282 million cases and 610,000 deaths were reported across 80 countries, with 95% of cases and deaths in Africa.The paper also cites projections of additional cases and deaths in Africa between 2024 and 2050 under current control levels.Source: §1 · p. 1
  • Microscopic Giemsa-stained blood smears are the gold standard for parasite identification and quantification, but are labor-intensive, time-consuming, expert-dependent, and unsuitable for resource-constrained settings.Source: §1 · pp. 1–2
  • RDTs are quicker but limited in species discrimination and sensitivity for low parasite densities and non-falciparum species, while PCR is costly, slow, and infrastructure-intensive.Source: §1 · p. 2
  • Prior deep-learning studies often used random cell-level splits that leak slides across training and test sets, while explainability evaluation was largely qualitative.The study frames its questions around architecture efficiency under slide-disjoint evaluation and the faithfulness and stability of Grad-CAM, SHAP, and LIME, including on misclassified samples.Source: §1 · p. 2

2. Related Work

Prior malaria-detection research reports strong CNN performance but also trade-offs in model size, latency, and accuracy. Work on explainability explores attribution methods, while pruning studies raise questions about preserving explanation quality.

2.1. Deep Learning for Malaria Detection

  • Compact CNNs and more complex attention-based models have both reported high NIH-dataset accuracy, while transfer-learning approaches reported 89%–95.7%.Reported examples include 99.51% accuracy for a four-layer CNN, 99.37% for a parallel CNN with soft attention, and 99.12% for an attention-based LSTM.Source: §2.1 · p. 2
  • Lightweight models can reduce computational demands but may trade away predictive performance, motivating compression approaches such as pruning.Prior pruning work reported compressing large CNNs by 30× to 49× with minimal accuracy loss.Source: §2.1 · pp. 2–3

2.2. Explainable AI in Malaria Detection

  • Malaria-detection XAI studies use Grad-CAM, SHAP, and LIME to highlight image regions or feature contributions, with reports of coarse spatial activation and finer pixel or superpixel attribution.Prior work also describes LIME as emphasizing cell-boundary detail, complemented by broader Grad-CAM activation patterns.Source: §2.2 · p. 3
  • Prior pruning studies report that aggressive weight removal can distort Grad-CAM explanations and surface clinically irrelevant concepts in a skin-lesion task.The cited skin-lesion example contrasts a model retaining 100% of weights with one pruned to 23%.Source: §2.2 · p. 3

3. Material and Methods

The study uses patient-grouped NIH malaria images to compare classification architectures, including a modified ResNet18 with test-time augmentation. It also applies three explainability methods to examine model predictions.

3.1. Dataset and Preprocessing

  • The NIH dataset contains 27,558 infected and uninfected cell images across 200 slides from 150 infected and 50 healthy patients, with expert annotations.The images come from Giemsa-stained thin blood smears, and image sizes are inconsistent.Source: §3.1 · p. 3
  • Images are grouped by slide identifier so cells from the same patient do not appear in multiple splits, making the evaluation slide-disjoint.The dataset includes variation in parasite morphology and artifacts such as dust and uneven staining; Figure 2 labels patient-level structure and overlap before versus after grouping.Source: §3.1 · p. 3, Figure 2 · p. 4
  • The 80/20 patient-grouped split uses 160 training slides and 40 test slides, with flipping, rotation, and color-jitter augmentation.Source: §3.1 · p. 4

3.2. Model Training and Evaluation

  • The proposed model uses a ResNet18 backbone with a custom classification head and is compared with MobileNetV2, EfficientNet-B2, ResNet101, VGG19, and ResNet18.The method describes a trainable convolutional backbone and test-time augmentation; Figure 3 is labeled as the ResNet18+TTA architecture.Source: §3.2 · p. 5, Figure 3 · p. 6

3.3. Explainability Measures

  • The best-performing model is examined with Grad-CAM for influential spatial regions, LIME for localized-perturbation effects, and SHAP for feature contributions.SHAP assigns pixel attribution scores to regions that positively or negatively influence the predicted class.Source: §3.3 · pp. 6–7

4. Results and Discussion

MobileNetV2 had the smallest model size and fastest inference, while the ResNet18+TTA ablation achieved the highest reported accuracy, F1-score, and AUC. Pruning reduced inference time and improved accuracy; XAI explanations aligned on correctly classified samples but diverged on a misclassified one.

4. Results and Discussion

  • MobileNetV2 had the smallest model size (8.49 MB) and fastest inference (1.35 ms/image), while ResNet18 achieved 96.92 % accuracy and was selected for the ablation.VGG19 had the largest model size (532.45 MB) and highest Recall (0.9898); EfficientNet-B2 achieved 97% accuracy at 2.13 ms inference speed.Source: §4 · p. 7, Table 1 · p. 7
  • 97.96 % accuracy, 0.9806 F1-Score, and 0.9966 AUC were reported for ResNet18 with TTA, a 0.02 pp gain over the baseline.The reported dominant effect of TTA was increased inference cost.Source: §4 · p. 7, Table 2 · p. 7
  • ResNet18+TTA performed well from the beginning in validation accuracy despite higher training loss, which the authors interpret as strong generalization and resistance to overfitting.The training-loss and validation-accuracy plots are shown in Figures 5 and 4, respectively.Source: §4 · p. 7, Figure 4 · p. 8, Figure 5 · p. 8
  • The confusion matrix reports 2,675 Uninfected and 2,440 Parasitized predictions, with 71 Uninfected and 51 Parasitized samples misclassified.The sample predictions shown in Figure 7 correctly classified all four examples.Source: §4 · p. 8, Figure 6 · p. 9, Figure 7 · p. 9
  • 30% pruning reduced inference time across five TTA augmentations from 14.55 ms to 10.65 ms and improved accuracy by an additional 0.21%.The unpruned baseline in this comparison was 97.90%, 0.06 pp below the Table 2 result; the authors describe that difference as within the test-set noise floor.Source: §4 · p. 9, Figure 8 · p. 10
  • SHAP, LIME, and Grad-CAM attributions aligned on the two correctly classified samples, while the final Parasitized sample was misclassified as Uninfected.The paper reports differing attribution locations across methods for that misclassified case and suggests deeper diagnostic investigation.Source: §4 · p. 10, Figure 9 · p. 11

5. Conclusions

The study benchmarks malaria classifiers using a group-aware split, proposes modified ResNet18+TTA, and examines model explanations. ResNet18+TTA led in accuracy and AUC, while pruning improved its accuracy and inference time; explanations diverged on misclassified cases.

  • A group-aware split kept files sharing an identifier prefix together, and the study benchmarked five CNN architectures alongside modified ResNet18+TTA with XAI.Source: §5 · p. 11
  • 97.96% accuracy and 0.996 AUC made ResNet18+TTA the strongest reported model, while pruning reached 98.11% accuracy and cut inference to 10.65 ms.Source: §5 · p. 11
  • Explanations agreed on correctly classified cases but diverged on misclassifications, which the authors identify as a manual-review red flag; cross-dataset validation and expert-annotation faithfulness remain future work.Source: §5 · p. 12

Related papers

Loading 2609.31682v2…