AICOM-MP: an AI-based monkeypox detector for resource-constrained environments

Abstract

Under the Autonomous Mobile Clinics (AMCs) initiative, the AI Clinics on Mobile (AICOM) project is developing, open sourcing, and standardising health AI technologies on low-end mobile devices to enable health-care access in least-developed countries (LDCs). As the first step, we introduce AICOM-MP, an AI-based monkeypox detector specially aiming for handling images taken from resource-constrained devices. We have developed AICOM-MP with the following principles: minimisation of gender, racial, and age bias; ability to conduct binary classification without over-relying on computing power; capacity to produce accurate results irrespective of images' background, resolution, and quality. AICOM-MP has achieved state-of-the-art (SOTA) performance. We have hosted AICOM-MP as a web service to allow universal access to monkeypox screening technology, and open-sourced both the source code and the dataset of AICOM-MP to allow health AI professionals to integrate AICOM-MP into their services.

Keywords:

• Health AI
• monkeypox
• autonomous mobile clinics
• SDG3

1. Introduction

As illustrated in Figure 1, AMCs are self-driving vehicles equipped with medical diagnostic equipment, telemedicine capability and an artificial intelligence (AI) software that can perform some of the tasks of health professionals, such as disease screening and basic diagnostics. Such clinics have the potential to revolutionise health-care delivery by bringing health-care services to hard-to-reach populations (Liu, Kong, et al., 2022).

Figure 1. Enabling technologies for Autonomous Mobile Clinics.

Source: Created by Shaoshan Liu, Permission Granted

Particularly, AMCs answer to the United Nations (UN) Sustainable Development Goal 3 (SDG3) on health, which represents a universal recognition that health is fundamental to human capital and social and economic development (The United Nations, n.d.). Despite the progress achieved at the global level, many health systems are not sufficiently prepared to respond to the needs of the rapidly ageing population. Inequitable access to healthcare continues to impede progress towards achieving universal health coverage (UHC). AMCs empower the provision of essential health services that are safe, affordable, and effective, with a special emphasis on the poor, vulnerable and marginalised segments of the population (World Economic Forum, n.d.).

While AMCs serve as mobility platforms for essential health services provisioning, the health AI engines incorporated onto AMCs determine the reach and coverage of AMCs. In essence, health AI engines are applications on AMCs, similar to how various applications expand the usage scenarios of smart phones. This way, AMCs should be configured with different medical devices and health AI engines as the deployment situation demands (Guan et al., 2021; Liu, Huang, et al., 2022).

Under the AMC initiative and facing the recent global monkeypox outbreak, in this article we introduce AICOM-MP, an open-source AI-based monkeypox detector that has achieved SOTA performance and capable of handling images taken from resource-constrained devices, such as low-end mobile phones, computing environments often endured by people in LDCs (International Telecommunication Union, n.d.; Liu, Yu, et al., 2022).

Through this project, we have made the following contributions to the field of health AI:

• We have constructed the AICOM-MP dataset to optimise coverage, diversification, and generalisation, and we have open-sourced the dataset for other health AI professionals to improve upon (Section 3).

• We have developed AICOM-MP in the form of a web service to allow universal access of monkeypox screening technology. We have open-sourced AICOM-MP source code for other health AI professional to integrate AICOM-MP into their services (Section 4).

• We have demonstrated that AICOM-MP, trained on AICOM-MP dataset, effectively filters out noises from real-world data and achieves SOTA detection results. Particularly, the restoration unit within the AICOM-MP pipeline effectively handles images taken on low-end devices (Section 5).

• With the methodology introduced in this article, we have generalised a methodology for developing health AI engines to enrich the coverage of AMCs (Section 6).

2. Monkeypox symptoms and screening methods

Monkeypox is a highly contagious skin disease which causes rash-like lesions (World Health Organization, n.d.). Up to 24 October 2022, there have been 75,568 global cases in 109 locations, 74,677 of which occurred in locations that did not report any monkeypox cases historically  (U.S. Center of Disease Control, n.d.). The lesion regions can be seen on several body parts including hands, faces, genital areas, and normally turn into scabs before healing. Patients may experience symptoms such as fever, swollen lymph nodes, and this can give rise to various medical complications such as sepsis.

Clinical diagnoses for Monkeypox can be conducted through a biopsy or a polymerase chain reaction (PCR) laboratory test  (World Health Organization, n.d.). However, these tests are expensive, and people in LDCs may not be able to afford mass-scale PCR testing. In contrast, utilising AI for monkeypox screening is cost efficient and can be deployed in a mass-scale. Multiple research groups (Nafisa Ali et al., 2022; Patel et al., 2022; Sitaula & Shahi, 2022) have trained monkeypox classification machine learning models. In addition to monkeypox, AI-powered medical screening technologies have demonstrated notable efficacy in accurately classifying COVID-19 symptoms manifested in pulmonary images (A. Gupta et al., 2020; Kedia & Katarya, 2021; Shyni & Chitra, 2022). They highlight the potential of AI in mitigating the impact of global pandemics by assisting public health decision-making.

Different from other existing works, we have developed AICOM-MP according to the following principles: minimisation of gender, racial, and age bias; ability to conduct binary classification when the image resolution/quality is low; capacity to produce accurate results irrespective of images' background. We have carefully constructed an AICOM-MP training dataset to meet the above requirements. The AICOM-MP dataset is also balanced so th at the amount of monkeypox images and “others” (i.e. other diseases, healthy, etc.) images are approximately equal. Our approach has achieved SOTA performance for AI-based monkeypox detection.

3. Monkeypox dataset for model training

Datasets are key to machine learning model training, in this section we review existing monkeypox datasets, summarise their shortcomings, and explain how the AICOM-MP Dataset that we developed addresses these problems.

3.1. Existing monkeypox datasets

We start by performing a comprehensive survey of existing monkeypox datasets as follows:

• Monkeypox Skin Image Dataset 2022: Islam et al. (2022) populated this Monkeypox dataset through web-scraping Monkeypox, Chickenpox, Smallpox, Cowpox, and Measles infected skin as well as healthy-skin images. This dataset contains 804 original web-scrapped images, including 160 Monkeypox images, 178 chickenpox images, 358 smallpox images, 54 cowpox images, 47 measles images, and 50 healthy-skin images. These images have been sequentially screened by two expert physicians specialised in infectious diseases to validate the supposed infection.

• Monkeypox-dataset-2022: Ahsan et al. (2022) generated this dataset also through web-scraping from websites, newspapers, and online portals and publicly shared samples using internet search engines. This dataset contains 171 original web-scrapped images, including 43 monkeypox images, 47 chickenpox images, 17 measles images, and 54 healthy-skin images. Seven augmentation operations have been applied to the web-scrapped images using ImageDataGenerator from Keras.

• Monkeypox Skin Lesion Dataset: Nafisa Ali et al. (2022) developed this dataset through manual web-scrapping from public case reports, news portals, and websites. Skin lesion images were verified using Google's Reverse Image Search and cross-referenced with other sources. The repeated, out-of-focus, low-resolution, and low-quality images were discarded using a two-stage screening process. The dataset contains 228 original web-scrapped images, including 102 monkeypox images, and 126 others (i.e. chickenpox, measles) images. Thirteen augmentation operations were applied to the web-scrapped images.

As displayed in Table 1, despite different approaches of organising and augmenting image data have been applied in each dataset, shortage of high-quality monkeypox lesion images and imbalance of images from torso, race, and age groups remain to be the major problems. Most of the monkeypox images have low resolutions, and some of them are artificially generated. These raise our concern that the machine learning model may treat certain resolution level as a sign of monkeypox images.

Table 1. Specifics of existing monkeypox datasets.

Dataset	Collection Method	Validation Method	Augmentation Operations	Monkeypox Images	Chickenpox Images	Smallpox Images	Cowpox Images	Measles Images	Healthy-skin Images	Others (i.e. chickenpox, measles)	Total
Monkeypox Skin Image Dataset 2022	Web-scraping	Expert physicians specialised in infectious diseases	NA	160	178	358	54	47	50	NA	804
Monkeypox -dataset-2022	Web-scraping	NA	7	43	47	NA	NA	17	54	NA	171
Monkeypox Skin Lesion Dataset	Web-scrapping	Google's Reverse Image Search, cross-reference, two-stage screening process	13	102	NA	NA	NA	NA	NA	126	228

3.2. AICOM-MP dataset

As illustrated in Figure 2, we utilised  Nafisa Ali et al. (2022) as the foundation dataset, then we enriched and balanced the dataset by including more unique monkeypox images that exhibit salient monkeypox lesion features, over-sampling the dataset to include equal number of healthy-skin images from different torso, gender, racial, and age groups, and using carefully selected techniques to augment images from the minority class (monkeypox) until reaching a balanced state with respect to the majority class (others) among the dataset. To improve the model's generalisation ability such that it can accurately map the noisy and transformed images to an accurate categorisation, we have trained the AICOM-MP model on the augmented dataset but used the original un-augmented images for validation.

Figure 2. AICOM-MP dataset.

Source: Created by Tianze Yang, Tianyi Yang, Permission Granted

3.2.1. Data selection

On the basis of Nafisa Ali et al. (2022), we have added 18 healthy-skin images and 8 monkeypox images from Sitaula and Shahi (2022). Additionally, we selected 18 healthy hand and 18 healthy face images from the 11k hands dataset  (Afifi, 2019) and Flickr-Faces-HQ Dataset (Karras et al., 2019). We collected 22 newly found monkeypox images from 7 published articles whose subject matter is monkeypox disease (Bruno et al., 2022; Girometti et al., 2022; Kabuga & El Zowalaty, 2019; McCollum & Damon, 2014; Meaney-Delman et al., 2022; Minhaj et al., 2022; Petersen et al., 2019). As a result, the AICOM-MP dataset contains 312 original images, of which 132 images are classified to monkeypox and 180 images are classified into others. We used this dataset as our validation/testing dataset which was split by a ratio of 65/35. To verify how well our trained model performs under real-world settings, we generated a new test dataset COCO_MP dataset, which consists of 200 manually selected COCO 2017 dataset (Lin et al., 2014) and 132 original monkeypox images. The diversity of sources of the AICOM-MP dataset ensures the coverage of our trained models.

3.2.2. Data augmentation

Data augmentations increase the coverage, diversification, and the generalisation ability of training dataset. We executed augmentations based on image geometric manipulations and did not incorporate photometric transformations for quality control purposes, as photometrically permutated images makes little sense from a human perspective (Shorten & Khoshgoftaar, 2019). To augment the newly appended images, we followed the techniques applied by the precedent dataset (Nafisa Ali et al., 2022).

3.2.3. Data balancing

Data balancing acts as a tool against impartial favour towards the majority class(es) during models' internal decision process. After applying such a technique, we generated 1818 augmented monkeypox images and 2466 augmented “others” images incorporated into the training dataset. We balanced the training dataset by augmenting 618 monkeypox images using the ImageGenerator function from Tensorflow Keras. The aforementioned validation/testing dataset was also balanced by augmenting 48 monkeypox images using the same function.

4. AICOM-MP monkeypox detector architecture

In this section, we delve into the AICOM-MP architecture. Note that, we have implemented a web service to host AICOM-MP, which is able to efficiently process images taken on a wide spectrum of mobile devices. Particularly, we have optimised AICOM-MP to process images taken on resource-limited devices, such that people from LDCs can utilise the AICOM-MP technology.

The AICOM-MP pipeline resembles how professionals diagnose pox symptoms through visual inspections. We model such expertise into the medical AI and perform medical diagnoses based on an expert information retrieval manner. In particular, the image processing step operates similarly to physicians' visual behaviours of allocating their visual attention towards lesion regions. AICOM-MP's model architecture is illustrated in Figure 3.

Figure 3. AICOM-MP model architecture.

Source: Created by Tianze Yang, Permission Granted

4.1. Human-object segmentation/background removal

As the first component of the analytical pipeline, this “level-1” vision layer determines the object towards which the vision model should allocate its attention, letting the model know “what” to look for. It plays the role of channelling attention in ViT and constitutes the object selection process. We selected the $U^2$-Net DL model and used its pre-trained weights to carry out background removal tasks (Qin et al., 2020). Its two-level nested U-structure architecture promises a computational-efficient yet powerful performance.

By removing non-targeted components from the image space, $U^2$-net helps increase model's robustness in dealing with images taken in noisy and complex real-world environments and reduces the chances for the model to misdirect its attention to irrelevant salient objects. Additionally, if a human object occupies the majority of space, $U^2$-net would view the object itself as background and remove a large proportion of the object. Based on empirical observation, a blacked-out region that occupies more than 87% of the original image space would output meaningless results that have no use for medical inferences, and we decided to preserve the image without applying segmentation in that case.

4.2. Region-based skin detection and segmentation

To further leverage the reduced spatial information from inputted images, this “level-2” vision layer plays the role of enhancing discrimination against non-skin features. It constitutes the adaptive spatial region selection mechanism of spatial attention in ViT and determines “where” the model should focus.

We selected an open-sourced FCNResNet10 model and used its pre-trained weights to carry out region-based skin segmentation task (Brennan, 2022). Consistent with the previous vision layer, a blacked-out mask that covers more than 87% of the original image space causes significant loss of information that is medically valuable, and would be excluded from being applied to the inputted images.

Based on our empirical evaluations, we noticed improvements brought by blocking regions covered by non-skin objects (i.e. rings, sweaters, etc.) from the model's inputting field. Following this discovery, we proceeded to testing the necessity of former layer as to whether there is an advantage of combining these two attention mechanisms. The ablation experiments reveal notable improvements brought by combining the two attention mechanisms. Collaboratively, the former level helps prevent the latter from being misled and removing valuable regions from the image space.

4.3. Deep-learning (DL) based classification

As the last component of the pipeline, this “level-3” vision layer is designed to allocate the model's eventual attention towards skin lesions across the reduced image space. It constitutes ViT's branch attention and determines “which” the model is looking for. It is assigned with a classification task of detecting monkeypox symptoms. To enable the use of DL model in energy-constraint devices while maintaining robust monkeypox screening performances, we resort to EfficientNetB7, which is based on MobileNet and has an architecture designed for performance on mobile CPU (Tan & Le, 2019). Furthermore, in light of the value demonstrated by MobileNet in COVID-19 screening tasks  (A. Gupta et al., 2020), it becomes evident that the family of EfficientNet models also exhibit the potential to address other medical-image diagnostic challenges effectively.

To ensure the model's performance in a limited yet domain-specific dataset, we did a transfer learning on the pre-trained ImageNet weights (Deng et al., 2009). Inspired by model training methods described in  Kunwar (n.d.), four additional layers were attached to the EfficientNetB7 model: a batch normalisation layer where momentum = 0.99 and $ϵ=0.001$; a dense layer containing a l2 kernel regulariser with l $=0.016$, l1 activity regulariser and l1 bias regulariser with l $=0.006$, and relu activation function; a dropout layer whose dropout rate $=0.45$; an output layer with softmax activation function to produce binary classification results.

The end-to-end processing pipeline of AICOM-MP is illustrated in Figure 4. During the implementation process, we tuned the parameters as follows: each image was resised into 224*224 dimensions; adaptive momentum estimation (Adam) was used as the optimiser; batch size was set at 48; momentum was set at 0.99; and the initial learning rate was set at 0.001. To prevent over-fitting, we utilise the learning rate decay over each epoch coupled with the kernel and activity regularisers.

Figure 4. AICOM-MP workflow.

Source: Created by Tianyi Yang, Permission Granted

4.4. Resolution restoration

One of AICOM-MP's objectives is to handle pictures taken on low-end phones, these pictures usually have low image resolution. To achieve this, we have implemented a resolution restoration unit in the AICOM-MP pipeline. But to avoid altering the perception of monkeypox characteristics by the machine learning model, we use images of original resolution (without enhancing) when training the model. An ablation study that investigated the effectiveness of this component is given in Section 5.

5. Evaluation

In this section, we perform a comprehensive study to evaluate AICOM-MP. We compare AICOM-MP to existing models on various datasets as well as real-world data to demonstrate that AICOM-MP is the SOTA in monkeypox screening.

Specifically on the detection accuracy results, Ali et al. have presented a performance comparison of different deep-learning models (VGG 16, ResNet50, InceptionV3) and an ensemble model. The four models respectively achieved detection accuracy ranging from 74% to 90% on the binary monkeypox classification dataset proposed by Nafisa Ali et al. (2022). As presented in Table 2, AICOM-MP has achieved 100% accuracy on Ali et al.'s binary monkeypox classification tasks.

Table 2. Model performance statistics on Ali et al. Dataset (Nafisa Ali et al., 2022).

Validation: Testing scores on Ali et al. Monkeypox Dataset (Nafisa Ali et al., 2022)
	Precision (%)	Recall (%)	F1 (%)	Accuracy (%)
AICOM-MP${}_3$ (AICOM-MP dataset)	100 : 100	100 : 100	100 : 100	100 : 100

To verify that AICOM-MP is not over-fitting to achieve perfect scores, we evaluated AICOM-MP on our experimental dataset consisted of 6124 images, in which 2596 images belong to the monkeypox class and 3528 images belong to the other class, augmented using different parameter settings than the training augmentation settings. Note that, our focus is not solely on improving the classification accuracy of our model, but also on improving the existing datasets by considering fairness issues (torso/gender/racial/age/bias), discovering related methods to balance datasets, and designing an analytical pipeline suitable for medical practices and real-world challenges. Conclusion 1: As shown in Table 3, AICOM-MP outperforms other proposed methods, and has achieved SOTA results on various monkeypox datasets.

Table 3. Model performance statistics on various datasets.

Performance comparison on the experimental dataset
Model	Weighted-Pre.	Weighted-Recall	Weighted-F1
Adjusted EfficientNetB3 trained on Nafisa Ali et al. (2022) dataset	0.8673	0.8601	0.8608
AICOM-MP trained on  Nafisa Ali et al. (2022) dataset	0.9102	0.8827	0.8844
AICOM-MP trained on  Nafisa Ali et al. (2022) dataset + healthy-skin images	0.9295	0.9136	0.9147
AICOM-MP trained on AICOM-MP dataset	0.9650	0.9634	0.9635

Next we evaluated AICOM-MP on real-world data, obtained from COCO_MP dataset as described in Section 3. Table 3 implies that not only the selection of DL models is of the utmost importance, but the quality of training datasets greatly influences the models' performance. Specifically, when we switched from using EfficientNetB3 to using EfficientNetB7 while keeping  (Nafisa Ali et al., 2022) as their common training dataset, we noticed an increase of at least 2.26% and at most 4.29% across the models' precision, recall, and F1 score; when we switched from using  (Nafisa Ali et al., 2022) to using AICOM-MP dataset to train a EfficientNetB7 model, the model's accuracy, recall, and F1 score all increased by at least 5.48% and at most 8.07%. We also observe such effect in Table 4, where we demonstrated that the EfficientNetB7 trained using AICOM-MP dataset achieved an accuracy of 88.55% on COCO_MP dataset whereas the EfficientNetB7 trained using (Nafisa Ali et al., 2022) had an accuracy of 40.66%. In comparison, AICOM-MP has achieved the highest generalisability in terms of not giving misdiagnoses when handling real-world images, and highest capability of capturing subtle and salient characteristics of monkeypox lesion among noisy environments. Conclusion 2: AICOM-MP has achieved SOTA monkeypox detection results on real-world data.

Table 4. Ablation over the COCO_MP dataset using proposed architecture.

Ablation study results
AICOM-MP (Nafisa Ali et al., 2022 dataset)	AICOM-MP (AICOM-MP dataset)	Restoration	Background removal	Skin segmentation	Accuracy (%)
•					40.66
	•				88.55
	•	•			91.57
	•		•		95.18
	•			•	95.48
	•	•	•		96.08
	•		•	•	96.99
	•	•		•	95.18
	•	•	•	•	96.34

Finally, we analyse how each technique in the AICOM-MP pipeline improves detection performance, and the results have been summarised in Table 4. Comparing to the original AICOM-MP (Nafisa Ali et al., 2022 dataset) model, AICOM-MP (AICOM-MP dataset) model itself increases the accuracy by more than two folds, suggesting not only the accurate representation of the medical information embedded within the dataset, but also its extensive generalisation ability to real-world environments. Thus, AICOM-MP (AICOM-MP dataset) exhibited high monkeypox screening performances and it constitutes level-3 vision layer. Additionally, using $U^2$-Net to suppress background noises, the accuracy achieved by level-3 vision layer was increased by 6.63%, demonstrating a rectifying effect $U^2$-Net has on level-3 vision layer. In short, $U^2$-Net has proven to be useful in improving level-3 vision layer's monkeypox screening performance and it constitutes level-1 vision layer. Nonetheless, many features from human gadgets may appear as indications of monkeypox infections, thereafter inducing incorrect diagnoses. We resorted to a skin segmentation algorithm to construct the level-2 vision layer to prune non-skin features, and successfully increasing the model's accuracy by 6.93%. We obtained the most optimised performance by further combining level-1, level-2, and level-3 vision layers. The new architecture minimises the effect that real-world distractions have on the level-3 vision layer as it boosts the model's accuracy to 96.99%, demonstrating the effectiveness of superposing level-2 vision layer on level-1 vision layer and the efficacy of our designed attention allocation pipeline in medical-image processing. Conclusion 3: the AICOM-MP pipeline trained on AICOM-MP dataset effectively filters out noises from real-world data and achieves SOTA detection results.

In addition, we analyse the effectiveness of the proposed restoration unit using COCO_MP dataset. The ablation results shown in the second and third rows of Table 4 indicate an increase in classification accuracy by 3.02% when the restoration unit was used before passing into AICOM-MP. Conclusion 4: the restoration unit within the AICOM-MP pipeline effectively handles images taken on low-end devices.

6. Discussion

In this section, we summarise the learnings from the AICOM-MP project and generalise a methodology that can be applied to other health AI “applications” beyond monkeypox, such as measles, chickenpox, and eczema.

• Robust Dataset: A robust dataset that has wide coverage, diversification, and generalisation is essential to improve detection accuracy of a health AI engine. In the case of AICOM-MP, We have developed a dataset through data selection to improve coverage, data augmentation to improve diversification, and data balancing to improve generalisation.

• Attention Model: Through this project, we have identified that a model that resembles how medical professionals diagnose diseases using the attention mechanism performs the best. The image processing step operates similarly to physicians' visual behaviours on allocating their visual attention towards lesion regions. Images are then processed hierarchically through three stages of stratification of visual information, and only the medically indicative features are kept for further inferences. This attention-based method can be generalised to other health AI engines as well.

• Compute Optimisation: Compute optimisation is imperative to ensure efficiency and scalability of health AI engines so that these engines can serve unlimited patients. In AICOM-MP, we have optimised both data processing and handling images from resource-constrained devices. For data processing, we have incorporated a compressor component in our web service. We have also utilised an image restoration module to improve detection accuracy on pictures taken from low-end mobile devices.

7. Conclusion

This article is the first to explore the development of health AI applications for AMCs, which are mobility platforms for essential health services provisioning, especially for people with limited health-care access. We deem AMCs as the next generation of health-care delivery platforms, whereas health AI engines are applications on these platforms, similar to how various applications expand the usage scenarios of smart phones. We chose monkeypox as our starting point as the world is currently under a global monkeypox outbreak, and the technology developed can be of immediate help. Not only did we achieve SOTA monkeypox detection results, we have hosted the technology on a web service to allow universal access, and we have open sourced the code and dataset for health AI professionals to integrate into their services. Most importantly, the methodology developed in this article can be generalised to many other diseases, leading to the development of many health AI engines for various AMCs usage scenarios. As an immediate next step, we will explore integrating blockchain mechanisms to further improve patient data privacy on the AICOM platform (B. B. Gupta et al., 2021; Sun et al., 2023).

Disclosure statement

No potential conflict of interest was reported by the author(s).

Data availability statement

The data and source code that support the findings of this study are available at https://github.com/Tim-Yang-YTY/AICOM.