[Paper Review] Malaria Detection and Classificaiton

Ruskin Raj Manku, Ayush Sharma+1 more|arXiv (Cornell University)|Nov 29, 2020

Digital Imaging for Blood Diseases11 references7 citations

TL;DR

This paper proposes a two-stage deep learning framework for malaria detection and classification using thin blood smear images, combining Faster R-CNN for infected cell detection and a fine-tuned ResNet-50 with a 4-layer classifier for stage-level classification. The method achieves 82% accuracy on a publicly available field-stained dataset, establishing a benchmark for future research in low-resource settings.

ABSTRACT

Malaria is a disease of global concern according to the World Health Organization. Billions of people in the world are at risk of Malaria today. Microscopy is considered the gold standard for Malaria diagnosis. Microscopic assessment of blood samples requires the need of trained professionals who at times are not available in rural areas where Malaria is a problem. Full automation of Malaria diagnosis is a challenging task. In this work, we put forward a framework for diagnosis of malaria. We adopt a two layer approach, where we detect infected cells using a Faster-RCNN in the first layer, crop them out, and feed the cropped cells to a seperate neural network for classification. The proposed methodology was tested on an openly available dataset, this will serve as a baseline for the future methods as currently there is no common dataset on which results are reported for Malaria Diagnosis.

Motivation & Objective

• To develop an automated, field-deployable malaria diagnosis system suitable for low-income regions with limited access to trained pathologists.
• To address the lack of a standardized, publicly available dataset for fair comparison of malaria detection methods.
• To improve classification accuracy by decoupling detection and classification into two specialized stages, reducing feature loss from pooling layers.
• To validate the model on field-stained (inexpensive) blood smears, not just high-quality giemsa-stained images, for real-world applicability.
• To provide a reproducible baseline using an openly accessible dataset to support future method development and evaluation.

Proposed method

• A two-stage approach is employed: first, Faster R-CNN detects infected red blood cells (RBCs) by classifying them as 'infected' (encompassing ring, trophozoite, schizont, gametocyte stages) and generating region proposals.
• The detected infected regions are cropped and resized to a fixed size to preserve image features, avoiding information loss from RoI pooling in end-to-end Faster R-CNN.
• The cropped infected cell images are then passed through a pre-trained ResNet-50 model to extract 2048-dimensional feature vectors.
• These feature vectors are fed into a 4-layer fully connected neural network for stage-level classification into four parasite stages: ring, trophozoite, schizont, and gametocyte.
• The dataset is balanced with 140 images per class (560 total), with 90% used for training and 10% for testing to ensure balanced evaluation.
• The model is trained on a publicly available dataset of thin blood smears stained with field stain, an inexpensive alternative to giemsa, enhancing field applicability.

Experimental results

Research questions

• RQ1 Can a two-stage deep learning pipeline outperform a single-stage Faster R-CNN in classifying malaria parasite stages while preserving image features?
• RQ2 How does performance on field-stained blood smears compare to high-quality giemsa-stained images in terms of detection and classification accuracy?
• RQ3 Can a publicly available, field-relevant dataset serve as a reliable benchmark for future malaria diagnosis models?
• RQ4 Does decoupling detection and classification reduce feature loss from RoI pooling and improve classification accuracy?
• RQ5 What level of precision, recall, and accuracy is required for reliable deployment in low-resource, field-level diagnostic settings?

Key findings

• The two-stage approach achieved an overall accuracy of 82% on the test set, with macro F1-score of 0.82 and weighted F1-score of 0.82.
• The model achieved a mean average precision (mAP) of 0.832 and recall of 0.69 in the detection phase, indicating strong performance in identifying infected cells.
• The classification model showed high precision and recall: 0.91 for ring stages, 0.87 for schizonts, 0.80 for gametocytes, and 0.73 for trophozoites.
• When compared to a one-stage Faster R-CNN that directly classifies parasite stages, the two-stage method achieved significantly better results, with mAP increasing from 0.4 to 0.832.
• The study establishes the first publicly available benchmark dataset for malaria diagnosis using field-stained thin blood smears, enabling fair comparison across future methods.
• The model demonstrates robustness to field-stain images, which are more common in low-income regions, suggesting practical utility in real-world diagnostic settings.