Source-linked AI summary
Monkeypox Skin Lesion Detection Using Deep Learning Models: A Feasibility Study
Shams Nafisa Ali, Md. Tazuddin Ahmed, Joydip Paul, Tasnim Jahan, S. M. Sakeef Sani, Nawsabah Noor, Taufiq Hasan
TL;DR
Early monkeypox diagnosis is difficult because its skin lesions resemble those of chickenpox and measles, while PCR confirmation may be unavailable and public datasets are lacking. The study introduces the MSLD dataset and evaluates transfer-learning models with augmented images and 3-fold cross-validation; ResNet50 achieves the best accuracy, and the best model is deployed in a prototype web application.
Problem
Early monkeypox diagnosis is challenging because its clinical features resemble chickenpox and measles, PCR testing is not widely available, and no public monkeypox skin-lesion dataset exists.
Method
The study introduces the openly accessible MSLD dataset and evaluates VGG16, ResNet50, InceptionV3, and their ensemble using transfer learning, data augmentation, and 3-fold cross-validation.
Results
82.96±4.57% accuracy is achieved by ResNet50, outperforming VGG16 at 81.48±6.87% and the ensemble at 79.26±1.05%; the best model is deployed in a prototype web application.
Takeaways & Limitations
The preliminary results indicate potential for AI-assisted early monkeypox detection and remotely deployable screening tools when traditional testing is unavailable.
Takeaways & Limitations
The limited number of unique patients reduces model generalization capability across geographical, racial, and gender variability.
Abstract
from arXiv · showhide
The recent monkeypox outbreak has become a public health concern due to its rapid spread in more than 40 countries outside Africa. Clinical diagnosis of monkeypox in an early stage is challenging due to its similarity with chickenpox and measles. In cases where the confirmatory Polymerase Chain Reaction (PCR) tests are not readily available, computer-assisted detection of monkeypox lesions could be beneficial for surveillance and rapid identification of suspected cases. Deep learning methods have been found effective in the automated detection of skin lesions, provided that sufficient training examples are available. However, as of now, such datasets are not available for the monkeypox disease. In the current study, we first develop the ``Monkeypox Skin Lesion Dataset (MSLD)" consisting skin lesion images of monkeypox, chickenpox, and measles. The images are mainly collected from websites, news portals, and publicly accessible case reports. Data augmentation is used to increase the sample size, and a 3-fold cross-validation experiment is set up. In the next step, several pre-trained deep learning models, namely, VGG-16, ResNet50, and InceptionV3 are employed to classify monkeypox and other diseases. An ensemble of the three models is also developed. ResNet50 achieves the best overall accuracy of $82.96(\pm4.57\%)$, while VGG16 and the ensemble system achieved accuracies of $81.48(\pm6.87\%)$ and $79.26(\pm1.05\%)$, respectively. A prototype web-application is also developed as an online monkeypox screening tool. While the initial results on this limited dataset are promising, a larger demographically diverse dataset is required to further enhance the generalizability of these models.
I. INTRODUCTION
Monkeypox is difficult to diagnose early because its skin manifestations resemble other diseases and confirmatory PCR testing may be unavailable. The study addresses this gap by introducing a dataset, evaluating deep-learning models, and developing a prototype screening application.
- Monkeypox closely resembles chickenpox and measles, making early diagnosis challenging when skin-rash differences are subtle.
- A prototype web application incorporates the developed models to assess uploaded skin-lesion images for potential monkeypox infection.
- PCR confirmation is not widely available, while early detection, contact tracing, and isolation are important for limiting community transmission.
- Deep-learning systems can extract image features automatically, but their effectiveness depends on large, suitable datasets that remain difficult to obtain for monkeypox.
- The study introduces the openly accessible Monkeypox Skin Lesion Dataset containing monkeypox, measles, and chickenpox images collected from public sources.
II. BACKGROUND
Monkeypox is a globally disseminated zoonotic disease with clinical features overlapping other rash illnesses. When laboratory confirmation is unavailable, skin-lesion assessment and AI-based tools may support earlier identification.
- Monkeypox is a zoonotic Orthopoxvirus disease historically associated with tropical rainforests of west and central Africa.
- By 30 June 2022, 5135 laboratory-confirmed cases had been reported from 66 member states across multiple WHO regions.
- Monkeypox was assessed by WHO as a moderate global health risk after spreading in non-endemic countries without epidemiological connections to endemic areas.
- Symptoms typically last 2 to 4 weeks, with an incubation period of 5 to 21 days and a rash appearing within 1 to 3 days of fever onset.
- Histopathology, virus isolation, and PCR support diagnosis, while skin-lesion examination and AI systems may assist when these tools are unavailable.
III. DATASET PREPARATION
The MSLD was assembled from publicly available images of monkeypox and similar diseases, quality-screened and standardized for binary classification. Augmentation expanded the scarce original dataset substantially.
- The dataset focuses on binary classification of monkeypox versus an Others class containing chickenpox and measles images.
- Images were manually collected from case reports, news portals, and websites, then verified, screened for quality, cropped, and resized to 224 × 224 pixels.
- The original dataset contains 228 images: 102 monkeypox images and 126 Others images.
- The sample images are organized with Monkeypox examples in the left panel and measles or chickenpox examples in the right panel.
A. Pre-trained Models
The study evaluates three ImageNet-pre-trained CNN architectures using transfer learning for monkeypox skin-lesion classification. The accompanying augmentation illustration shows the range of transformations applied to training images.
- A. Pre-trained Models: VGG16, ResNet50, and InceptionV3 were selected as ImageNet-pre-trained CNN architectures for transfer-learning-based classification.
- A. Pre-trained Models: VGG16 uses stacked 3 × 3 convolutional filters and factorized convolution to support deeper feature extraction while limiting overfitting.
- A. Pre-trained Models: ResNet uses residual modules in which convolution is followed by Batch Normalization and ReLU non-linearity.
- A. Pre-trained Models: The augmentation example includes rotations, translation, reflection, shear, color jitter, noise, blur, and scaling transformations.
B. Implementation Details
The selected pre-trained networks were adapted for lesion classification by partially unfreezing their backbones and adding fully connected layers with dropout, using Adam optimization and binary cross-entropy.
- B. Implementation Details: Input images were resized to (224, 224, 3), and the bottom eight layers of each pre-trained backbone were unfrozen.The models then used three fully connected layers with 4096, 1072, and 256 nodes, followed by dropout.
- B. Implementation Details: The networks were implemented in Keras and accelerated with Nvidia K80 GPUs using batch size 16.Training used Adam with an initial learning rate of 10^-5 and binary cross-entropy loss.
C. Experimental Design
The study used patient-independent 3-fold cross-validation with approximate 70:10:20 training, validation, and test splits, augmenting only training and validation images.
- C. Experimental Design: The original images were divided into training, validation, and test sets in approximate proportions of 70:10:20 while maintaining patient independence.This partitioning was used for the three-fold cross-validation experiment.
- C. Experimental Design: Data augmentation was applied to training and validation images, whereas the test set contained only original images.The split and augmentation strategy is illustrated in Fig. 5.
- C. Experimental Design: Accuracy, precision, F1-score, and sensitivity were used as performance metrics.
D. Results
ResNet50 achieved the strongest accuracy in the 3-fold evaluation, while the best-performing model was deployed in a prototype web application for initial lesion assessment.
- D. Results: 82.96±4.57% accuracy was achieved by ResNet50, exceeding VGG16 at 81.48 ± 6.87%.The ensemble did not outperform ResNet50, although it had the lowest accuracy standard deviation across folds.
- D. Results: The ensemble system showed the most consistent accuracy across the three folds because it had the lowest accuracy standard deviation.
- D. Results: The best-performing model was deployed in a prototype web application for monkeypox screening.Users can upload a skin-lesion photograph and receive an initial assessment.
- D. Results: Table II presents the performance comparison of the different deep learning models and the proposed ensemble.
V. DISCUSSION
The study presents promising preliminary findings, but limited patient diversity, ImageNet pre-training, and missing clinical metadata constrain wider applicability and generalizability.
- V. DISCUSSION: The limited number of unique patients reduces the models’ generalization capability.The authors suggest larger samples with better geographical, racial, and gender distribution.
- V. DISCUSSION: ImageNet pre-training may limit performance because ImageNet contains no skin lesion images.The authors propose multi-source dermatoscopic data for pre-training to improve accuracy and generalizability.
- V. DISCUSSION: Web-scraped images lack diagnostic metadata such as clinical history, disease-onset timing, and disease stage.The authors call for international collaboration to collect a larger, demographically generalizable dataset.
VI. CONCLUSION
The study presents MSLD and evaluates transfer-learning models for monkeypox lesion detection, finding promising feasibility despite the dataset’s small size.
- MSLD is an open-source dataset for automatic monkeypox detection from skin lesions.
- The feasibility study evaluates VGG16, ResNet50, and InceptionV3 using transfer learning and 3-fold cross-validation.
- The authors suggest that MSLD and the prototype web application could support remotely deployable screening and early detection when traditional testing is unavailable.