Skin diseases are a common ailment affecting millions of people worldwide.The surface of the skin has many different features that are affected by the environment, dust, smoke, harmfulmaterials and othersuch factors. These effects occuras per the type of skin- Oily skin or Dry skin. Early detection and treatment of these diseases are crucial in improving patient outcomes. In recent years, machine learning techniques, particularly convolutional neural networks (CNN), have shown promise in the automated detection of skin diseases.
In this paper, after analyzing such harmful effects, andtry to classify the type of skin disease, so as to provide further treatment. AbetterConvolutionalNeuralNetwork- EfficientNet B0 is used to predict the type of skin disease based on image samples of skin diseases provided. The proposed system consists of three main stages: preprocessing, training, and testing. In the preprocessing stage, the images were re-sized, normalized, and augmented to enhance the quality of the images. In the training stage,theEfficientNetB0CNNarchitecturewasusedtotrainthe systemontheskindiseaseimages.Thesystem\'sperformancewas evaluated using a test dataset, which included images that were notusedinthetrainingprocess.Thisgivesanimprovedefficiency over other existing models.
Introduction
The skin is the largest human organ and various skin conditions—like psoriasis, eczema, and acne—affect millions worldwide. Skin type (dry, oily, acne-prone) significantly influences susceptibility to these disorders. Accurate diagnosis is critical but often time-consuming and subjective.
Machine learning (ML), especially deep learning using convolutional neural networks (CNNs), can help automate skin type classification and disease detection by analyzing large datasets of skin images. EfficientNet B0, a state-of-the-art CNN architecture, shows promise due to its accuracy and efficiency.
A literature survey reviewed previous research on plant disease detection and skin cancer classification using CNNs, highlighting the importance of dataset selection, feature extraction, network depth, and fine-tuning.
The motivation is to develop an automated skin disease detection system using EfficientNet B0, improving diagnosis speed and accuracy, reducing dermatologists' workload, and enabling early treatment.
Challenges include limited diverse datasets, class imbalance, computational demands, and clinical deployment concerns.
The proposed work involves training EfficientNet B0, fine-tuning it by freezing early layers, and using an improved YOLOv3 algorithm for feature extraction and detection. The model was implemented in Python using PyTorch and trained on a curated dataset of seven skin disease categories, achieving 95.5% accuracy.
Conclusion
Inthisstudy,examinedandevaluateddifferent versions of the EfficientNet neural network for imageclassificationonasmalldataset.Itisfound that EfficientNet B0 provided a good balance between accuracy and computational efficiency, making it a promising option for training and classifying on small datasets with limited computational resources.
Despite providing slightly less accuracy compared to other EfficientNet architectures, we wereabletocompensate foritthroughourimage selection and preprocessing process. This suggests that EfficientNet B0 can learn from the limited data available and generalize well to unseen examples. Moreover, able to further improve the accuracy from 85.5% to 95.5% within just five epochs, demonstrating the potentialofEfficientNet B0fortraining onsmall datasets.
Also explored the use of the YOLOv3 algorithm for fine-tuning the features, which furtherimprovedtheaccuracy ofourmodel.Ourfindings suggest that the combination of EfficientNet B0 and YOLOv3 provides a powerful and efficient solution for image classification tasks.Forourfuturework,itisplannedtoupscalethe model and increase the size of our dataset to further improve the accuracy. We aim to implement a larger dataset with a significantly increased accuracy, which can help us further validate the effectiveness of EfficientNet B0 in this context.
Overall,ourfindingshighlight theimportance of choosing an appropriate neural network architecture for image classification tasks, especiallywhen working with smalldatasets and limited computational resources.The use of EfficientNet B0 and YOLOv3 provides an effective solution that balances accuracy and efficiency, making it a promising option for various applications
References
[1] Denesh Kumar et. al., “EfficientNet based Convolutional Neural Network for Visual Plant Disease Detection”, 978-1-6654- 85845/22.
[2] Tan Nhat Pham et. al., “Early Disease Classification of Mango Leaves Using FeedForward Neural Network and Hybrid Metaheuristic Feature Selection”, 10.1109/ACCESS.2020.3031914.
[3] Karen Simonyan, Andrew Zisserman, “Very Deep Convolutional Networks for LargeScaleImageRecognition”, arXiv:1409.1556.
[4] Mingxing Tan, Quoc V. Le, “EfficientNet: Rethinking model scaling for convolutional neural networks”, arXiv:1905.11946v5 [cs.LG] 11 Sep 2020.
[5] Jayme Garcia ArnalBarbedo, “A review on the main challenges in automatic plant disease identification based on visible range images”, 10.1016/j.neuri.2021.100034.
[6] K.Aliet.al.,“MulticlassskincancerclassificationusingEfficientNets–afirststep towards preventing skin cancer”, 10.1016/j.biosystemseng.2016.01.017.
[7] H. Hong, et.al., “Tomato Disease Detection and Classification by Deep Learning”, 10.1109/ICBAIE49996.2020.00012.
[8] Y. Ai, et. al., “Research on Recognition Model of Crop Diseases and Insect Pests Based on Deep Learning in HarshEnvironments”, 10.1109/ACCESS.2020.3025325
[9] J.Deng,et.al.“ImageNet:Alarge-scalehierarchical image database” 10.1109/CVPR.2009.5206848
[10] S.Verma,M.A.Razzaque,U.Sangtongdee,C. Arpnikanondt, B. Tassaneetrithep, and A. Hossain, “Digital diagnosis of Hand, Foot, and mouth disease using hybrid deep neural networks,”IEEEAccess,vol.9,pp.143481– 143494, 2021.
[11] F.W.Alsaade,T.H.H.Aldhyani,andM.H. Al-Adhaileh, “Developing a recognition systemfordiagnosingmelanomaskinlesions using artificial intelligence algorithms,” Computational and Mathematical Methods in Medicine, vol. 2021, Article ID 9998379, 20 pages, 2021.
[12] X. Zhu, X. Li, K. Ong et al., “Hybrid AIassistive diagnostic model permits rapid TBS classification of cervical liquid-based thin-layer cell smears,” Nature Communications,vol.12,no.1,pp.3541–3612,2021.
[13] A.M.Malciu,M.Lupu,andV.M.Voiculescu, “Artificial intelligence-based approaches to reflectance confocal microscopyimageanalysis inDermatology,” JournalofClinicalMedicine,vol.11,no.2,p.429,2022.
[14] J. Divakaran, S. Malipatil, T. Zaid et al., “Technicalstudyon5Gusingsoftcomputing methods,” ScientificProgramming,vol.2022.