Source-linked AI summary
Image Data collection and implementation of deep learning-based model in detecting Monkeypox disease using modified VGG16
Md Manjurul Ahsan, Muhammad Ramiz Uddin, Mithila Farjana, Ahmed Nazmus Sakib, Khondhaker Al Momin, Shahana Akter Luna
TL;DR
Monkeypox image-data scarcity limits work on image-based diagnosis. This paper presents a publicly available dataset and evaluates a modified VGG16 model in two studies, reporting accuracy from 78% to 97%.
Problem
The lack of publicly available Monkeypox image data limits researchers' ability to develop image-based diagnosis models.
Source: §2.1 · p. 3, §5 · p. 12Method
The authors construct a publicly available image dataset and evaluate a modified VGG16 model in two studies, using LIME to explain predictions.
Source: §5 · p. 12Results
The modified VGG16 distinguished patients with Monkeypox symptoms from others in both studies, with reported accuracy ranging from 78% to 97%.
Source: §5 · p. 12Takeaways & Limitations
The dataset is publicly available for experiments and commercial use, and the authors use LIME to explain model predictions.
Source: §5 · p. 12The dataset contains limited samples due to time constraints and is collected primarily from open sources rather than a hospital or clinical facility.
Source: §4 · p. 12
Abstract
from arXiv · showhide
While the world is still attempting to recover from the damage caused by the broad spread of COVID-19, the Monkeypox virus poses a new threat of becoming a global pandemic. Although the Monkeypox virus itself is not deadly and contagious as COVID-19, still every day, new patients case has been reported from many nations. Therefore, it will be no surprise if the world ever faces another global pandemic due to the lack of proper precautious steps. Recently, Machine learning (ML) has demonstrated huge potential in image-based diagnoses such as cancer detection, tumor cell identification, and COVID-19 patient detection. Therefore, a similar application can be adopted to diagnose the Monkeypox-related disease as it infected the human skin, which image can be acquired and further used in diagnosing the disease. Considering this opportunity, in this work, we introduce a newly developed "Monkeypox2022" dataset that is publicly available to use and can be obtained from our shared GitHub repository. The dataset is created by collecting images from multiple open-source and online portals that do not impose any restrictions on use, even for commercial purposes, hence giving a safer path to use and disseminate such data when constructing and deploying any type of ML model. Further, we propose and evaluate a modified VGG16 model, which includes two distinct studies: Study One and Two. Our exploratory computational results indicate that our suggested model can identify Monkeypox patients with an accuracy of $97\pm1.8\%$ (AUC=97.2) and $88\pm0.8\%$ (AUC=0.867) for Study One and Two, respectively. Additionally, we explain our model's prediction and feature extraction utilizing Local Interpretable Model-Agnostic Explanations (LIME) help to a deeper insight into specific features that characterize the onset of the Monkeypox virus.
1 Introduction
The introduction presents Monkeypox as a growing concern and identifies missing public image data and limited image-based ML research as barriers. It proposes a public image dataset, modified VGG16 detection, and LIME explanations.
1 Introduction
- Cases rose from 50 in 1990 to 5,000 in 2020, and 2022 reports extended beyond Africa as Monkeypox transmission can also occur between people.The introduction describes transmission through contact with infected people, animals, or materials, and identifies skin bumps among longer-term effects.Source: §1 · p. 2
- The introduction describes skin-lesion observation and exposure history as initial diagnostic steps, with electron microscopy or PCR used for definitive confirmation.It also reports no appropriate treatment under CDC guidelines, while two oral drugs previously used for smallpox had been used for Monkeypox and vaccination was described as the ultimate solution.Source: §1 · p. 2
- Prior ML studies addressed other medical-imaging and skin-disease tasks, but the authors had found no study of image-processing-based Monkeypox diagnosis.Examples in the cited literature include deep-learning COVID-19 diagnosis and image-based skin-disease classification.Source: §1 · pp. 2–3
- The authors identify two barriers: no publicly accessible Monkeypox dataset for training and testing, and the need for time to develop a suitable algorithm and model.The second barrier is linked to the virus's recent spread across many nations.Source: §1 · p. 3
- The authors argue that a Monkeypox image dataset could let researchers begin AI-assisted work; their starting set had 164 samples and reached 1,915 after augmentation.They cite an earlier COVID-19 image collection as inspiration and anticipate their dataset serving a similar research role.Source: §1 · p. 3
- The stated contributions are a publicly available Monkeypox image dataset, a modified VGG16 detector, and LIME-based post-image explanations of model predictions.The authors say the LIME explanations offer insight into features characterizing Monkeypox onset.Source: §1 · p. 3
2 Methodology
The methodology builds and evaluates a modified VGG16 using collected and augmented skin images, with separate studies, LIME-based superpixel analysis, and standard classification metrics. Data are divided into training and test sets, and Study One results are summarized with a 95% confidence interval.
2.1 Data collection
- Images were gathered from online sources, with non-Monkeypox examples including Chickenpox, Measles, and normal images, some collected from consenting participants.Normal images were also collected manually from participants who reported no skin-disease symptoms, with consent forms used for approval.Source: §2.1 · pp. 3–4
2.2 Data augmentation
- The datasets were expanded with Keras ImageDataGenerator using image transformations, and the augmentation procedure resized images to 128 × 128 and generated batches of 16.The pseudocode loops from 1 to 20; listed transformation options include rotation, width and height shifting, and flipping.Source: §2.2 · pp. 4–5
2.3 Transfer learning approaches
- The modified VGG16 combines a pre-trained feature extractor with added convolutional, pooling, flattening, dense, and dropout layers.The described input size is 224 × 224 pixels.Source: §2.3 · p. 5
- Study One tuning selected a batch size of 10, learning rate of 0.001, and 35 epochs, while Study Two explored larger batch sizes and epoch counts.The Study Two search ranges were batch sizes 30, 50, 70, and 100, and 50, 70, 100, and 150 epochs.Source: §2.3 · p. 6
2.4 LIME as explainable AI
- LIME parameters were used to calculate superpixel values for image analysis.The text describes superpixels as regions produced by over-segmenting an image and used to help identify image features.Source: §2.4 · p. 7
2.5 Experiment setup
- Study One used 43 Monkeypox and 47 Chickenpox images, while Study Two used 587 augmented Monkeypox and 1,167 Other samples, with an 80/20 train/test split.Other combines Chickenpox, Measles, and Normal images; five experimental runs were averaged for the reported outcome.Source: §2.5 · pp. 7–8
2.6 Performance evaluation
- Performance assessment included accuracy, precision, recall, F1-score, sensitivity, specificity, and AUC; Study One results were represented with a 95% confidence interval.The F1-score is described as the harmonic mean of precision and recall.Source: §2.6 · p. 8
3 Results
Study One performed better than Study Two in the reported preliminary results, while the figures show differences in epoch-wise performance, misclassification, AUC, and LIME explanations.
3 Results
- Study One significantly outperformed Study Two overall; its train-set accuracy was up to 9% higher, although both studies showed high sensitivity and low specificity on test data.The results are reported with 95% confidence intervals because the dataset contains few samples.Source: Table 5 · p. 9, §3 · p. 9
- Study One reached its highest accuracy and began overfitting after 35 epochs, while Study Two peaked at 100 epochs on training and validation data.Training and validation loss in Study One decreased through 35 epochs.Source: Figure 4 · p. 9, §3 · p. 9
- AUC was 0.972 on Study One’s train set and 0.748 on Study Two’s test set, the best and worst reported performances, respectively.Source: §3 · p. 9
- Train-set misclassification was 2.7% for Study One and 12.33% for Study Two; the authors suggest Study Two’s 1:1.98 class imbalance as a possible reason.Source: §3 · p. 9, Figure 5 · p. 10
- LIME showed that a Study One image with Monkeypox symptoms was misclassified as Chickenpox while the model focused mainly on an area without Monkeypox symptoms.The proposed model correctly classified all four Study Two test images shown.Source: §3 · p. 10
4 Discussions
The discussion reports results from two modified-VGG16 studies and describes LIME-based analysis, while noting limited samples and open-source collection as study constraints.
4 Discussions
- 0.83 ± 0.085 accuracy on a very small dataset and 0.78 ± 0.022 test-set accuracy on an imbalanced dataset.The authors report that no prior study was available for direct performance comparison because no Monkeypox image dataset existed.Source: §4 · p. 10
- LIME analysis was used to inspect top image features and showed models learning from and localizing infected regions.The feature figure covers classification among Monkeypox, Chickenpox, and Others; doctors also analyzed the data collection and model performance.Source: §4 · p. 10, Figure 7 · p. 11
- Time constraints left the developed dataset with limited samples.Source: §4 · p. 12
- The data were collected primarily from open sources rather than a hospital or clinical facility.Source: §4 · p. 12
5 Conclusion
The conclusion presents a public Monkeypox image dataset and a modified VGG16 evaluated in two studies, reporting accuracy from 78% to 97% and LIME explanations of predictions.
5 Conclusion
- 78% to 97% accuracy was reported for the modified VGG16 across two studies distinguishing Monkeypox symptoms from other cases.Source: §5 · p. 12
- The publicly available image dataset addresses Monkeypox image-data scarcity and can be used without privacy restrictions, including for commercial purposes.Source: §5 · p. 12
- LIME was used to explain model predictions, and doctors crosschecked those predictions.Source: §5 · p. 12
- The authors identify continued image collection, evaluation on highly imbalanced data, future comparisons, and mobile-tool deployment as further work.Source: §5 · p. 12