Towards Field-Ready AI-based Malaria Diagnosis: A Continual Learning Approach

Louise Guillon, Soheib Biga, Yendoube E. Kantchire, Mouhamadou Lamine Sane, Grégoire Pasquier, Kossi Yakpa, Stéphane E. Sossou, Marc Thellier, Laurent Bonnardot, Laurence Lachaud, Renaud Piarroux, Ameyo M. Dorkenoo · 2025

Read Towards Field-Ready AI-based Malaria Diagnosis: A Continual Learning Approach by Louise Guillon, Soheib Biga, Yendoube E. Kantchire, Mouhamadou Lamine Sane, Grégoire Pasquier, Kossi Yakpa, Stéphane E. Sossou, Marc Thellier, Laurent Bonnardot, Laurence Lachaud, Renaud Piarroux, Ameyo M. Dorkenoo (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2507-23648.txt.

Introduction

Louise Guillon and co-authors adapt a malaria diagnosis tool for field use in a 2025 paper. Malaria hits low-resource areas hard, where expert microscope readers may be scarce. Deep learning tools read thin blood smear images well, but they can fail at new sites with unlike conditions. Few practical fixes exist.

The team treats the problem as learning that must grow domain by domain. A YOLO detector adapts to new sites while keeping skill on old ones. Four continual learning strategies are tested on a multi-site clinical data set. Rehearsal based methods, which replay old examples, help most, they report. Field ready tools may follow.

Catalog record

CREATORLouise Guillon, Soheib Biga, Yendoube E. Kantchire, Mouhamadou Lamine Sane, Grégoire Pasquier, Kossi Yakpa, Stéphane E. Sossou, Marc Thellier, Laurent Bonnardot, Laurence Lachaud, Renaud Piarroux, Ameyo M. Dorkenoo
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE18 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2507-23648

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Towards Field-Ready AI-based Malaria Diagnosis: A Continual Learning Approach Louise Guillon, Soheib Biga, Yendoube E. Kantchire, Mouhamadou Lamine Sane, Grégoire Pasquier, Kossi Yakpa, Stéphane E. Sossou, Marc Thellier, Laurent Bonnardot, Laurence Lachaud, Renaud Piarroux, Ameyo M. Dorkenoo Abstract Malaria remains a major global health challenge, particularly in low-resource settings where access to expert microscopy may be limited. Deep learning-based computer-aided diagnosis (CAD) systems have been developed and demonstrate promising performance on thin blood smear images. However, their clinical deployment may be hindered by limited generalization across sites with varying conditions. Yet very few practical solutions have been proposed. In this work, we investigate continual learning (CL) as a strategy to enhance the robustness of malaria CAD models to domain shifts. We frame the problem as a domain-incremental learning scenario, where a YOLO-based object detector must adapt to new acquisition sites while retaining performance on previously seen domains. We evaluate four CL strategies, two rehearsal-based and two regularization-based methods, on real-life conditions thanks to a multi-site clinical dataset of thin blood smear images. Our results suggest that CL, and rehearsal-based methods in particular, can significantly improve performance. These findings highlight the potential of continual learning to support the development of deployable, field-ready CAD tools for malaria. alaria, yolo, computer-aided diagnosis, generalization, object detection, thin blood smear images, continual learning 1 Introduction Malaria, a parasitic disease caused by Plasmodium , remains a major global health challenge. Despite important efforts for prevention, diagnostic and treatment, malaria accounted for close to 600,000 deaths in 2023 [ 1 ] . Early and accurate diagnosis is critical to mitigate its impact, with several diagnostic methods available. Rapid Diagnosis Tests (RDT) offer fast and reliable results, but they can suffer from false negatives due to parasite mutations, and false positives can occur after treatment, leading the WHO to recommend microscopic analyses of thick and thin blood smears as the gold standard. Thin smears, in particular, are used to identify Plasmodium within red blood cells (RBC). However, the analysis of thick and thin blood smears requires high levels of expertise, which may be unavailable in remote areas where urgent diagnosis are needed. To address this challenge, several works have proposed computer-aided diagnosis (CAD) tools for malaria, based on thin and thick smears [ 3 , 4 , 19 ] . High performance has been achieved at the patient level, with 100% sensitivity and 95% specificity for thick smears thanks to the Autoscope [ 3 ] . Using smartphone-acquired images, 91.8%, 92.5%, 91.1% were obtained at the patient level for accuracy, sensitivity and specificity respectively [ 23 ] . Good performances have also been obtained for thin smears, with accuracies of 98.62% and 98.44% reported at the cell and patient levels respectively [ 15 ] . Smartphone applications incorporating these algorithms are being developed and tested in the field [ 18 , 19 , 23 ] . While these works bring massive opportunities, the real-world deployment of such CAD systems remains hindered by a critical challenge: generalization. Indeed, for a CAD system to be used in the field, it must perform well across various clinical environments in spite of deep learning models often sensitive to such variations [ 24 ] . Malaria CAD is not spared from this challenge, as different clinical sites can exhibit significant domain shifts [ 3 , 8 , 5 ] . This is particularly critical for routine clinical data which present challenges such as imbalanced dataset and varying image quality due to differences in equipment, staining and operators, making the task even more challenging. To address the site effect and improve generalization, strategies such as fine-tuning or joint training, can be used [ 5 ] . Continual learning (CL) offers another promising approach, where the model is trained incrementally, adapting to new domains or tasks while retaining knowledge from previous ones. CL can be categorized into task, domain, and class incremental scenarios [ 13 ] . In the context of this article, we are dealing with a domain-incremental scenario where the same task, detecting Plasmodium , is performed across different sites. Here, we propose the first application of continual learning for malaria CAD. We focus specifically on routine thin blood smear images and investigate how CL applied to YOLO models can help overcome generalization, paving the way for more robust, field-ready diagnostic tools. 2 Related works 2.1 Malaria Computer-Aided Diagnosis Two main methods have been proposed for thin blood smear images: one-step approaches, which treat the problem as an object detection task, and two-step approaches, which first extract the red blood cells before classifying them. In one-step methods, the model is trained to detect parasites on images of a thin smear for example. In this setting, models like Faster RCNN [ 6 , 18 ] , the YOLO family [ 22 , 12 , 15 ] and RT-DETR [ 4 ] have been used. However, the clinical applicability of such malaria CAD systems highly depends on their generalization capabilities across various environments. Although generalization has been largely underexplored in malaria diagnosis, studies have shown that performance can degrade significantly when CAD systems are applied to data from different clinical sites [ 5 , 8 ] , with accuracy dropping by over 20% at the image level in some cases [ 5 ] . These results highlight the need for approaches that can enhance generalization ability of CAD systems, and continual learning could offer a promising solution. 2.2 Continual Learning Continual learning is a deep learning paradigm aiming to address generalization by enabling models to adapt and accumulate knowledge over time, without forgetting previously learned information, commonly known as catastrophic forgetting . CL balances between two challenges: plasticity, which allows the model to learn new tasks, and stability, which helps retain knowledge from previous tasks [ 25 ] . Three main strategies have been introduced: rehearsal, regularization and architectural. The rehearsal strategy involves storing samples from previous tasks in a memory buffer and reusing them during training to prevent forgetting. Regularization-based methods, such as Elastic Weight Consolidation (EWC) [ 10 ] and Learning Without Forgetting (LWF) [ 14 ] , introduce regularization terms in the loss to preserve important parameters and avoid catastrophic forgetting. Architectural methods isolate task-specific parameters. CL has become an important area of research in medical imaging [ 13 ] , but to our knowledge, it has not yet been studied in the context of malaria diagnosis and very few studies have implemented CL for YOLO models [ 17 ] . 3 Methods Problem Formulation. We define the problem as a domain-incremental scenario where the goal is to train a model that can effectively adapt to new domains while minimizing catastrophic forgetting. Let T={T_{1},T_{2},...,T_{T}} represent the sequence of tasks. Each task involves a dataset D_{t} consisting of thin smear images, and the corresponding labels indicating infected RBC. We seek to train models, f_{t} , that learn from the sequence of tasks and must generalize well to each task T_{t} . Datasets. We consider five tasks (T=5), each corresponding to a dataset from a distinct site. The datasets include the publicly available NIH dataset 1 1 1 https://data.lhncbc.nlm.nih.gov/public/Malaria/NIH-NLM-ThinBloodSmearsPf/index.html , captured in Bangladesh [ 9 ] , and four in-house datasets collected from four

Read the complete text