Utilizing CNN Modeling as an Effective Tool for Skin Disease Differentiation and Detection of Flare-Ups

By Ethan Yu, USA

Abstract:Eczema is a global health concern that has impacted 15-30% of children and 2-10% of adults (Laughter et al., 2021). The illness is typically characterized by inflamed, red, and itchy skin. There is currently no cure for eczema and it is typically treated with a variety of medicines ranging from topical creams to oral medications. In many cases, the common characteristics often overlap with symptoms of other skin diseases such as acne and psoriasis, which cause emotional distress and discomfort that affect one's quality of life. This study aims to use Convolutional Neural Networks (CNN) to accurately detect eczema flare-ups and early symptoms. While other uses of machine learning to detect skin diseases have also been developed, they have either focused on a wider variety of skin diseases, used different CNN architectures, or focused only on a certain ethnic group. This approach can provide a more accurate prediction of early-moderate signs of eczema around the body thereby reducing the financial burden of diagnostics and the possibility of ineffective treatment stemming from misdiagnosis.

Introduction:

Eczema is a prevalent disease often found in younger children and causes a variety of unpleasant sensations. The disease is associated with skin barrier defects and immune dysregulation. It specifically involves two main groups of genes, namely mutations in the filaggrin (FLG) gene, which play a key role in the assembly of the stratum corneum. The underlying mechanisms associated with Atopic Dermatitis (AD), a specific type of eczema, such as IL-13, specifically induce AD through a Thymic stromal lymphopoietin (TSLP)-dependent pathway. Together with IL-22, these cytokines impact epidermal differentiation and reduce filaggrin production (Bantz et al., 2014). This cycle involving skin barrier defects that lead to immune system activation leads to a greater volume of inflammations and allergic sensitization, thereby laying the foundation for the Atopic March. Convolutional Neural Networks (CNNs) are a specialized neural network utilizing convolution to detect the spatial features of an image. By learning from a labeled dataset of eczema and non-eczema images, a CNN can differentiate between subtle visual cues that might not be easily noticeable to the human eye. This ability to detect nuanced features enables early diagnosis, allowing for timely intervention and treatment.

Literature Review:

With the rise of the prevalence of Eczema and other skin diseases around the world, coupled with the prominence of artificial intelligence (AI) and machine learning, many combined the two to improve diagnostics. Researchers have specifically developed a CNN model using the EfficientNet-b4 CNN algorithm and created a diagnosis assistant named AIDDA. The model was trained using a total of 4,740 clinical images to classify a variety of skin diseases ranging from. The model's overall accuracy was 95.80%±0.09%. However, the accuracy scores for the separate groups, psoriasis (87.62%), eczema, and AD (90.97%), were much lower compared to the healthy skin differentiation accuracy score of (98.62%) (Wu et al., 2020). In other cases, the use of CNNs has also been utilized in melanoma detection. Researchers used four different types of CNN architectures, ResNet, ResneXt, Se-ResNet, and SE-ResNeXt to classify malignant melanoma, melanocytic nevus, basal cell carcinoma, and actinic keratosis/Bowen’s disease. Combined with a softmax activation function and the Adam optimizer, they also utilized a weighted cross-entropy loss function to fix the weighting of class based on frequency.

Early stopping was also used to prevent overfitting. They were able to achieve a precision score of 0.85, 0.88, 0.84, and 0.82 respectively for each of their 4 models, ResNet, ResneXt, Se-ResNet, and SE-ResNeXt (Kwiatkowska et al., 2021). While these studies found success in the differentiation of multiple skin diseases, their accuracy in classifying the skin diseases as opposed to healthy skin failed to achieve an accuracy or precision score of 90% or more. Additionally, their datasets were filled with an overarching majority of one ethnicity. In contrast, this proposal aims to utilize a wide range of skin types found on open-source datasets. 

Methodology:​ 

Obtaining Dataset and Manual Annotation 

Obtain a large, diverse set of identified clinical images of human skin, including eczema-affected skin, acne-affected skin, psoriasis-affected skin, and healthy skin all from a variety of body parts including chest, face, arms, legs, elbows, and knees. These images should be sourced from public 

dermatology image databases (e.g., DermNet, Kaggle, or collected datasets with IRB approval). The dataset should include a wide variety of skin tones, and eczema severity levels, and then be annotated and separated into separate folders labeled as "eczema", "healthy.", and “other skin disease”. Finally, split the dataset into training (80%), and validation (20%). 

Creating CNN Model and Preprocessing 

Begin by importing Tensorflow, Numpy, Matplotlib, Seaborn, and Sklearn modules. Then begin the preprocessing stage by resizing all images to a standard dimension (e.g.,128x128 pixels) and converting them into TensorFlow datasets. To increase dataset diversity and reduce overfitting, apply data augmentation such as random horizontal flipping, slight rotations (±10%), and zooming (±10%). Normalize the pixel values from 0-255 range to 0-1. The first of the three convolutional layers should use 32 filters with a 3x3 kernel and ReLu activation for the identification of lower-level features such as edges. Then apply max pooling to reduce dimensionality. The second layer and third layer looks at more complex features such as textures, patterns, shapes, and coloring using 64 filters and 128 filters respectively, ReLu activation, and max pooling. After feature extraction, the output is flattened into a 1D vector (14x14x128) of 25,088 features. Then, a fully connected dense layer with 128 neurons applies ReLu activation to learn patterns, followed by a dropout layer with a rate of 0.5 to deactivate half the neurons and prevent overfitting. Finally, use an output layer with three neurons and a softmax function, a probability distribution for each of the three categories mentioned earlier. Once the architecture is established, complete the model using the categorical cross-entropy loss function to measure the difference between predicted class probabilities and labels, and the Adam optimizer to update weights based on frequency. Implement early stopping if validation loss does not improve for 5 consecutive epochs. Train for a maximum of 20 epochs with a batch size of 32. 

​

Statistical Analysis 

Generate a 3x3 confusion matrix to visualize the correctness of the classifications. Additionally, generate a classification report including accuracy, precision, recall, and f1 score, as well as a report on the misclassified images that represent model errors. Repeat the experiment for other skin diseases such as psoriasis and melanoma. This model can also be used in a mobile format for greater accessibility.

​

Conclusion:

Given the high mortality associated with late-stage lung cancer diagnosis in India, and the severe disparity in diagnostic access between urban and rural populations, there is an urgent need for scalable, low-cost early detection tools. By expanding my original research into a pan-India study, this project will enable the development of a robust, AIdriven, symptom-based lung cancer screening tool, adaptable for use in primary care settings across diverse regions. This tool can empower frontline healthcare workers to identify at-risk patients earlier, prioritize interventions, and ultimately save lives in communities where the traditional healthcare infrastructure is weak. By combining machine learning innovation with a commitment to healthcare equity, this research has the potential to meaningfully reduce the burden of lung cancer mortality across India.

​

References :

Bantz, S. K., Zhu, Z., Zheng, T. (2014). The atopic march: Progression from atopic dermatitis to allergic rhinitis and asthma. Journal of Clinical & Cellular Immunology, 5(2), 202. https://doi.org/10.4172/2155-9899.1000202 

​

Biagini Myers, J. M., & Khurana Hershey, G. K. (2010). Eczema in early life: Genetics, the skin barrier, and lessons learned from birth cohort studies. The Journal of Pediatrics, 157(5), 704-714. https://doi.org/10.1016/j.jpeds.2010.07.009 

​

Boguniewicz, M., & Leung, D. Y. M. (2011). Atopic dermatitis: A disease of altered skin barrier and immune dysregulation. Immunological Reviews, 242(1), 233-246. 

https://doi.org/10.1111/j.1600-065X.2011.01027.x 

​

Diagnosing eczema & dermatitis. (2025, February 27). NYU Langone Health. https://nyulangone.org/conditions/eczema-dermatitis/diagnosis 

Hill, D. A., & Spergel, J. M. (2018). The atopic march: Critical evidence and clinical relevance. Annals of Allergy, Asthma & Immunology, 120(2), 131-137. 

https://doi.org/10.1016/j.anai.2017.10.037 

​

International Eczema Council. (2022). Global report on atopic dermatitis 2022. https://www.eczemacouncil.org/assets/docs/global-report-on-atopic-dermatitis-2022.pdf Kwiatkowska, D., Kluska, P., & Reich, A. (2021). Convolutional neural networks for the detection of malignant melanoma in dermoscopy images. Postepy dermatologii i alergologii, 38(3), 412–420. https://doi.org/10.5114/ada.2021.107927 

​

Larsen, F. S., Holm, N. V., & Henningsen, K. (1986). Atopic dermatitis: A genetic-epidemiologic study in a population-based twin sample. Journal of the American Academy of Dermatology, 15(3), 487–494. 

Maulana, A., et al. (2023). Evaluation of atopic dermatitis severity using artificial intelligence. Narra J, 3(3), e511. https://doi.org/10.52225/narra.v3i3.511 

​

Mayo Clinic Staff. (2022, January 21). Atopic dermatitis (eczema) - symptoms and causes. Mayo Clinic. 

https://www.mayoclinic.org/diseases-conditions/atopic-dermatitis-eczema/symptoms-cau ses/syc-20353273 

​

Mayo Clinic Staff. (2022, January 21). Dermatitis (eczema) – diagnosis and treatment. Mayo Clinic. 

https://www.mayoclinic.org/diseases-conditions/dermatitis-eczema/diagnosis-treatment/d rc-20352386 

Shaw, T. E., et al. (2011). Eczema prevalence in the United States: Data from the 2003 National Survey of Children's Health. The Journal of Investigative Dermatology, 131(1), 67-73. https://doi.org/10.1038/jid.2010.251 

​

Smith Begolka, W., et al. (2021). Financial burden of atopic dermatitis out-of-pocket health care expenses in the United States. Dermatitis: Contact, Atopic, Occupational, Drug, 32(1S), S62-S70. https://doi.org/10.1097/DER.0000000000000715 

​

Wu, H., et al. (2020). A deep learning, image-based approach for automated diagnosis for inflammatory skin diseases. Annals of Translational Medicine, 8(9), 581. 

https://doi.org/10.21037/atm.2020.04.39 

​

Yotsu, R., et al. (2023, March 15). Deep learning for AI-based diagnosis of skin-related neglected tropical diseases: A pilot study. medRxiv: The Preprint Server for Health Sciences. https://doi.org/10.1101/2023.03.14.23287243