1.

Project title:
Development of skin disease detection software using RNN
2. Introduction / Background
Skin diseases are a major public health concern globally, affecting millions of individuals.
According to the World Health Organization (WHO), approximately 30% to 70% of the
population has experienced some form of skin disease during their lifetime. These conditions
not only affect physical health but can also lead to severe psychological issues, including anxiety
and depression. Skin diseases like eczema, psoriasis, and melanoma require early detection for
effective management and treatment. While many solutions use Convolutional Neural
Networks (CNNs) for image classification, there is potential in leveraging Recurrent Neural
Networks (RNNs) due to their capability to handle sequential data.
RNNs have demonstrated efficacy in tasks requiring pattern recognition and temporal
dependencies, making them suitable for analyzing skin images over time. This proposal outlines
developing a skin disease detection system that employs RNNs to classify and predict potential
skin conditions based on patient data and images.
3. Problem Statement
Skin diseases are a prevalent health issue in Africa and Uganda, with significant impacts on the
lives of individuals, families, and communities. In many regions, especially rural and
underserved areas, access to dermatological care is limited, which leads to delayed diagnosis
and treatment of skin conditions. According to the World Health Organization (WHO), Africa
bears a significant burden of neglected tropical diseases, many of which present with skin
manifestations such as ulcers, lesions, and rashes. These conditions often remain undiagnosed
or misdiagnosed due to the lack of skilled dermatologists and diagnostic tools, putting patients
at greater risk of complications, including infections and, in some cases, skin cancers like
melanoma.
In Uganda, dermatological services are mostly concentrated in urban areas, leaving the rural
population without easy access to specialized care. This lack of access leads to late-stage
diagnosis of severe conditions, making treatment more difficult and less effective. Skin diseases,
although often considered less life-threatening than other diseases, have profound effects on
quality of life, particularly for vulnerable populations. People suffering from skin conditions
frequently experience social stigma, psychological distress, and even economic hardship due to
their inability to work or participate in daily activities.
Furthermore, the tropical climate in Uganda and many parts of Africa exposes the population to
specific skin diseases such as fungal infections, dermatitis, and photo-induced skin conditions,
which are exacerbated by factors like limited hygiene facilities, exposure to the sun, and poor
living conditions.
Despite the high prevalence of skin diseases, there is a lack of awareness about the various
types of conditions and their potential severity. The general population often relies on
traditional medicine or self-diagnosis, leading to delays in seeking appropriate medical care.
This problem is compounded by the high cost of dermatological services and the shortage of
dermatologists in the country. The WHO reports that Uganda, like many other African nations,
faces a chronic shortage of healthcare workers, including dermatologists. As a result, skin
diseases often go untreated or are treated incorrectly, sometimes leading to severe
complications or even death in cases like advanced melanoma.
A skin disease detection system powered by Recurrent Neural Networks (RNN) could
significantly bridge these gaps, particularly in Uganda and other parts of Africa. This project
aims to provide solutions to the above challenges by developing an automated, affordable, and
accessible system that can detect and classify skin conditions based on images.
Therefore, this project addresses a critical healthcare gap in Uganda and Africa, where skin
disease is widespread but underdiagnosed, particularly in rural and low-resource settings. By
leveraging RNN technology, the project can contribute to early diagnosis, reduced healthcare
costs, and improved outcomes for those suffering from skin diseases.
4. Objectives
Main Objective:
 To develop a skin disease detection software using Recurrent Neural Networks (RNN) for
the early identification and classification of skin diseases.
Specific Objectives:
 To build a dataset of skin disease images and patient temporal data.
 To develop an RNN model capable of analyzing sequential skin images and detecting
disease progression.
 To design a user-friendly interface that allows users to upload images and receive
diagnostic insights.
 To evaluate the performance of the RNN model against existing CNN-based systems.
5. Justification
RNNs, known for processing sequential data, offer a unique advantage in tracking the
progression of skin diseases, which may evolve. This approach can contribute significantly to
the field of medical image analysis by providing enhanced diagnostic accuracy and personalized
treatment plans based on real-time progression patterns. Furthermore, this system will reduce
dependency on specialist consultations, enabling early diagnosis in remote and underserved
areas.
6. Significance of the Study
This research will:
 Improve early diagnosis and intervention in skin diseases.
 Increase accessibility to dermatological services in rural and remote areas.
 Provide a cost-effective alternative for skin disease monitoring.
 Offer healthcare providers additional tools for making informed diagnostic decisions
based on temporal data.
7. Scope of the Project
This project will focus on building an RNN-based skin disease detection system that accepts
patient image data over time to classify diseases and track their progression. The project will
primarily address common skin diseases such as eczema, psoriasis, and melanoma.
8. Literature Review
Skin disease diagnosis is a critical healthcare issue, with numerous research efforts focused on
developing automated solutions to improve early detection and treatment outcomes. Most of
the existing work uses Convolutional Neural Networks (CNNs) due to their success in image
classification tasks. However, there has been relatively less exploration of Recurrent Neural
Networks (RNNs) in this domain, even though RNNs could offer significant advantages in
tracking temporal patterns in diseases that progress over time.
This section will review the current literature on deep learning approaches to skin disease
detection, identify existing research gaps, and explain how this proposal aims to address these
gaps using a novel combination of CNNs and RNNs.
Current Research on Skin Disease Detection Using CNNs
CNNs have been widely used for skin disease classification, particularly for diagnosing skin
cancers such as melanoma, basal cell carcinoma, and squamous cell carcinoma. Esteva et al.
(2017) demonstrated that CNNs can achieve dermatologist-level accuracy in skin cancer
classification, using a dataset of over 129,000 clinical images across 2,032 different diseases.
Their research set a benchmark for automated dermatological diagnosis, showing that deep
learning systems could outperform even trained specialists in certain tasks.
Similarly, a study by Haenssle et al. (2018) compared CNN performance with that of
dermatologists in identifying malignant melanomas. The results indicated that the CNN
achieved higher sensitivity (86.6%) compared to dermatologists (74.4%), suggesting that AI-
based systems could significantly enhance diagnostic accuracy.
Limitations of CNNs:
While CNNs excel at image-based classification tasks, they are limited to analyzing static
images. This can be a significant drawback when diagnosing skin conditions that change or
progress over time. For instance, melanoma evolves through different stages, and early
detection is crucial. However, CNNs may not effectively capture the temporal evolution of a
skin lesion, as they treat each image independently. This is where RNNs, with their ability to
model sequential data, offer potential improvements.
Role of RNNs in Sequential Data Analysis
RNNs have been traditionally used for tasks involving sequential data, such as time-series
forecasting, speech recognition, and natural language processing. Long Short-Term Memory
(LSTM), a variant of RNNs, has been successful in solving problems where long-term
dependencies in the data must be captured (Hochreiter & Schmidhuber, 1997). LSTMs can
retain information across longer sequences, making them ideal for analyzing changes over time.
However, their application in the medical field, particularly in skin disease detection, is
relatively underexplored. Research by Liu et al. (2018) explored using RNNs in medical
diagnostics and applying them to electronic health records to predict patient outcomes. Though
not directly related to image processing, their work highlights the power of RNNs in making
predictions based on time-dependent information, such as the progression of diseases.
A key advantage of using RNNs, particularly LSTMs, in skin disease detection is their ability to
model temporal changes in skin conditions. This is especially useful for chronic diseases such as
psoriasis or evolving conditions like melanoma. By feeding a series of images taken over time
into an RNN, the model can learn how the disease progresses and detect early signs of serious
conditions based on patterns that might not be apparent from individual images.
Skin Disease Detection Using Hybrid Models
Recent work has begun to explore hybrid models that combine CNNs with RNNs to leverage the
strengths of both architectures. CNNs can be used to extract spatial features from images, while
RNNs (or LSTMs) can handle the temporal aspect, analyzing the evolution of these features over
time.
For example, the study by Tseng et al. (2017) used a CNN-RNN hybrid model to classify breast
cancer images based on both static features and temporal changes observed across multiple
mammograms taken over time. Their results showed that combining these approaches led to
better diagnostic accuracy than using CNNs alone. Although their work focused on a different
medical application, it demonstrates the potential of such hybrid models in handling temporal
changes in medical imaging.
Literature Gap:
Despite the potential of hybrid CNN-RNN models, there is limited research applying this
approach to skin disease detection. Most current systems treat skin disease diagnosis as a static
image classification problem. The temporal aspect, which could improve diagnostic accuracy
and early detection of conditions like melanoma, remains underutilized.
The Gap in Current Skin Disease Detection Systems
Based on the current literature, there is a significant gap in the use of deep learning models
that consider the temporal progression of skin diseases. Existing CNN-based models focus on
classifying skin conditions based on static images, ignoring how skin diseases evolve. Yet,
temporal changes in size, shape, or color are often crucial indicators for conditions like
melanoma. Moreover, many CNN-based systems, while accurate, often require large amounts
of labeled data to function effectively and are less effective for patients with rare or atypical
conditions where sequential imaging could provide additional diagnostic cues.
There are several gaps that this project aims to address:
Incorporating Temporal Data:
Existing systems do not adequately capture the progression of skin diseases over time.
This project proposes a hybrid CNN-RNN approach that can track temporal changes in
skin lesions, improving early detection and providing more comprehensive diagnostic
insights.
Personalized Disease Progression Monitoring:
Many current solutions do not provide a mechanism for monitoring the evolution of a
disease in an individual patient. The proposed system will allow users to upload
sequential images over time, offering personalized disease progression monitoring and
alerting them if significant changes occur.
Remote Accessibility:
Many deep learning-based skin detection systems are primarily research tools and are
not widely accessible to the public. This project aims to develop a user-friendly interface
that allows everyday users to upload images and receive diagnostic insights, thus
bridging the gap between research and real-world application.
Addressing the Gaps with CNN-RNN Hybrid Models in This Project
This project proposes a solution that addresses the gaps identified in the literature by using a
hybrid CNN-RNN model to analyze both spatial and temporal patterns in skin images. The
system will track the progression of skin diseases by feeding sequential images into an RNN,
allowing it to learn how skin conditions change over time.

Key Advantages:
Temporal Progression Analysis:
The RNN component will allow the system to recognize changes in skin lesions,
enhancing early detection for conditions like melanoma, where size and color changes
are crucial.
Feature Extraction with CNNs:
The CNN will extract spatial features from each image, such as color, texture, and lesion
boundaries. These features will be passed to the RNN for sequential analysis.
Personalized Tracking:
By analyzing images over time, the system will be able to provide personalized insights,
indicating whether a lesion is growing, shrinking, or changing shape. This is particularly
important for monitoring chronic conditions like psoriasis or tracking potentially
malignant changes in moles.
Comparative Evaluation:
The proposed system will be evaluated against existing CNN-based models to
demonstrate improvements in accuracy and early detection. The RNN’s ability to handle
sequential data is expected to outperform static-image-based CNN models, especially in
cases where temporal patterns are critical.
Conclusion of the Literature Review
Current research in skin disease detection has made significant strides using CNNs, but the
limitations in handling temporal data present an opportunity for innovation. By introducing
RNNs into the skin disease detection pipeline, this project aims to provide a more
comprehensive and accurate system capable of detecting and monitoring disease progression
over time. This hybrid approach will fill the gap in current research and offer significant
improvements in early detection, personalized care, and accessible diagnostic tools for skin
disease patients.
9. Methodology
The methodology is the most critical section of this proposal, as it outlines how the system will
be developed to meet the main objective (developing a skin detection software/system using
Recurrent Neural Networks for the early identification and classification of skin diseases) and
specific objectives. Each part of the methodology tackles one or more of the specific objectives,
ensuring a structured, objective-driven approach.
 Specific Objective Methodology Applied Tool(s) Used
- Online datasets (ISIC,
Data collection from public HAM10000).
datasets (e.g., ISIC, - Python (Pandas, NumPy)
1. To build a dataset of skin
HAM10000). for organizing data.
disease images and
temporal data
Collect sequential images for - Manual labeling or clinical
temporal analysis. partnerships for obtaining
time-series data.
Design and train a CNN-
2. To develop an RNN
RNN hybrid model. - TensorFlow/Keras or
model for analyzing
PyTorch for building and
sequential skin images and
CNN extracts spatial features, training the model.
detecting disease
RNN tracks temporal - Google Colab for free GPU.
progression
changes.
Build a web-based or - Flask or Django for the
standalone GUI for image backend.
upload and results. - HTML, CSS, and JavaScript
3. To design a user-friendly
for the front end.
interface for diagnostics
Display diagnostic results
with disease type and - OpenCV for preprocessing
progression. user-uploaded images.
4. To evaluate the - TensorFlow/Keras or
performance of the RNN Test the model using PyTorch for evaluation
model against existing validation and test datasets. metrics (e.g., accuracy,
CNN-based systems precision, recall).
Compare metrics like F1
- Matplotlib and Seaborn for
score and AUC-ROC with
visualization.
CNN-only systems.

Data Collection
(Objective 1: Build a dataset of skin disease images and temporal data from patients)
Data is at the core of any machine learning model, especially in a domain such as skin disease
detection. The data collection phase is designed to address Objective 1, which involves
gathering a robust dataset for training and evaluation purposes.
Skin Disease Image Dataset:
To train an RNN-based system effectively, the dataset must be large, diverse, and labeled
correctly. We'll focus on public datasets, such as the International Skin Imaging Collaboration
(ISIC) datasets and hospital records. These datasets contain high-quality images of various skin
conditions such as melanoma, eczema, and psoriasis. Each image will be associated with
metadata like patient ID, disease label, and possibly patient history.
Temporal Data:
Unlike static images, the proposed RNN model will track temporal changes in skin conditions.
To achieve this, multiple images from the same patient, taken at different time intervals
(weekly or monthly), will be collected. This approach ensures that the system learns how skin
diseases evolve over time. This is particularly useful for conditions like melanoma, where early
changes in size, shape, or color are crucial for diagnosis.
Data Preprocessing:
Preprocessing involves resizing images to a standard resolution, normalizing pixel values, and
augmenting the data. Augmentation (random rotations, flips, zooms) increases data diversity,
which is important for generalization. Noise reduction techniques will be used to remove
artifacts or irrelevant features in the images that could reduce the model’s performance.

Model Design
(Objective 2: Develop an RNN model capable of analyzing sequential skin images and
detecting disease progression)
Once the data is prepared, the heart of the project—building the RNN model—begins.
Objective 2 revolves around developing and training an RNN architecture for analyzing
temporal data (series of skin images over time).
Choosing the RNN Type:
Standard RNNs struggle with learning long-term dependencies, so we'll use a more advanced
variant, Long Short-Term Memory (LSTM) networks. LSTM can effectively remember and forget
information as required, making it ideal for tasks like skin disease progression detection, where
subtle changes over time matter.
Feature Extraction with CNN + RNN:
Before the data reaches the RNN layer, feature extraction will be done using a pretrained
Convolutional Neural Network (CNN). CNNs are excellent at learning spatial features from
images, like edges, textures, and patterns, which will help in identifying specific visual traits of
skin diseases. Once the CNN processes the image, the resulting features will be fed into the
LSTM network, which will then analyze the sequence of these feature vectors to track changes
in the skin condition.
Why CNN+RNN?:
This combination leverages the strengths of both networks: CNNs for spatial (image)
understanding and RNNs for sequential (time-based) understanding. The CNN extracts key
features from individual skin images, while the RNN analyzes the evolution of these features
over time.
Model Training:
The model will be trained using the Adam optimizer, with a suitable loss function such as
categorical cross-entropy. The training dataset will be split into training, validation, and test
sets to ensure proper evaluation and fine-tuning. Dropout techniques will be used to prevent
overfitting, especially since medical datasets can be prone to noise and data imbalance.
Disease Classification:
The output of the RNN will be a probability distribution across multiple skin diseases. The class
with the highest probability will be selected as the predicted disease, and if the disease is
cancerous, the system will also predict the stage based on temporal patterns (e.g., lesion size
increasing over time).
3. System Interface
(Objective 3: Design a user-friendly interface that allows users to upload images and receive
diagnostic insights)
Objective 3 involves creating a system that is not just accurate but also accessible and easy to
use. The interface will be web-based to maximize accessibility, particularly for users in remote
areas.
User Registration and Login:
Before accessing the diagnostic tool, users will need to create an account. This will allow
the system to keep track of uploaded images and provide continuous disease
monitoring over time.
Image Upload Feature:
After logging in, users can upload one or more images of their skin condition. The
uploaded images will be preprocessed to ensure they conform to the model’s input
requirements (e.g., resolution, file format).
Real-time Processing:
Once the image is uploaded, the system will run the preprocessed image through the
CNN+RNN pipeline. Depending on server resources, the model can either be hosted
locally (in a healthcare setting) or in the cloud (accessible via web interface).
Diagnostic Output:
Within a few seconds, the user will receive the predicted diagnosis. This will include:
 Disease Classification: The name of the detected skin condition (e.g., melanoma).
 Disease Progression: If temporal data is available, the system will indicate whether the
condition appears to be worsening or stabilizing.
 Confidence Score: A probability score indicating the confidence level of the system’s
prediction.
Additional Features:
 Doctor Recommendations: Based on the diagnosis, users will see a list of nearby
dermatologists for further consultation. The system will use geolocation APIs to identify
specialists.
 Feedback System: Users can provide feedback about the accuracy of the diagnosis or
report any errors, which will help in model improvement.
Model Evaluation
(Objective 4: Evaluate the performance of the RNN model against existing CNN-based
systems)
To ensure that the RNN-based system performs as expected, Objective 4 focuses on rigorous
evaluation and benchmarking.
Performance Metrics: The model will be evaluated using several key metrics:
 Accuracy: The percentage of correct predictions.
 Precision: The proportion of true positive predictions (correctly identified skin
conditions) out of all positive predictions.
 Recall: The proportion of true positives out of all actual positive cases.
 F1 Score: The harmonic mean of precision and recall, providing a balanced view of
performance.
AUC-ROC Curve: This metric will be used to evaluate the model’s ability to discriminate
between different skin conditions.
Comparative Analysis: We will compare the performance of the RNN-based system with
existing CNN-based systems. CNNs are generally strong at classifying images based on spatial
features, but they may struggle with temporal data (disease progression over time). The RNN’s
ability to handle sequential data should provide improved performance, especially in cases
where early-stage disease detection is critical.
Cross-validation: The model will undergo k-fold cross-validation to ensure that it generalizes
well to unseen data. This is crucial since the data set may not be perfectly balanced across all
skin conditions.
System Deployment
Once the model is trained, validated, and tested, it will be deployed as a web-based service.
The web interface will be hosted on cloud platforms such as Amazon Web Services (AWS) or
Google Cloud, ensuring scalability and accessibility.
Security Considerations: Given the sensitive nature of medical data, all user information and
uploaded images will be encrypted both at rest and in transit. The system will comply with
medical data protection standards, such as HIPAA in the U.S. and GDPR in Europe, ensuring that
patient privacy is safeguarded.
Maintenance and Continuous Learning: The system will continuously collect feedback from
users. This feedback will be used to fine-tune the model, ensuring that it becomes more
accurate over time. As new data is collected, the RNN can be retrained to improve its
predictions and adapt to evolving disease patterns.

10. Tools, Equipment, and Software

Programming Languages:
Python, TensorFlow, Keras
Software:
Jupyter Notebooks, PyTorch, OpenCV for image processing
Hardware:
High-performance GPU for model training
Dataset:
Publicly available skin disease datasets (e.g., ISIC 2019 Challenge Dataset)

REFERENCES:
Esteva, A., Kuprel, B., Novoa, R.A., Ko, J., Swetter, S.M., Blau, H.M., & Thrun, S. (2017).
Dermatologist-level classification of skin cancer with deep neural networks. Nature, 542(7639),
115–118. doi:10.1038/nature21056.
Haenssle, H.A., Fink, C., Toberer, F., Winkler, J., Stolz, W., Deinlein, T., ... & Hofmann-Wellenhof,
R. (2018). Man against machine: Diagnostic performance of a deep learning convolutional
neural network for dermoscopic melanoma recognition in comparison to 58 dermatologists.
Annals of Oncology, 29(8), 1836–1842. doi:10.1093/annonc/mdy166.
Hochreiter, S., & Schmidhuber, J. (1997). Long short-term memory. Neural Computation, 9(8),
1735–1780. doi:10.1162/neco.1997.9.8.1735.
Liu, C., Chen, Z., Xie, D., Zhang, L., & Zhang, Q. (2018). An end-to-end deep learning architecture
for electronic health record-based automatic mortality prediction. Scientific Reports, 8(1), 1–11.
doi:10.1038/s41598-018-24885-1.
Tseng, H., Luo, Y., Cui, S., Shi, X., & Wang, L. (2017). Hybrid deep learning framework for breast
cancer diagnosis using mammograms. Medical Image Analysis, 43, 176–185.
doi:10.1016/j.media.2017.10.004.

Esteva, A., Kuprel, B., Novoa, R.A., Ko, J., Swetter, S.M., Blau, H.M., & Thrun, S. (2017).
Dermatologist-level classification of skin cancer with deep neural networks. Nature, 542(7639),
115–118. doi:10.1038/nature21056.
Codella, N.C., Nguyen, Q.-B., Pankanti, S., Halpern, A., & Smith, J.R. (2017). Deep learning
ensembles for melanoma recognition in dermoscopy images. IBM Journal of Research and
Development, 61(4/5), 5-1. doi:10.1147/JRD.2017.2708299.
Hochreiter, S., & Schmidhuber, J. (1997). Long short-term memory. Neural Computation, 9(8),
1735–1780. doi:10.1162/neco.1997.9.8.1735.
Liu, C., Chen, Z., Xie, D., Zhang, L., & Zhang, Q. (2018). An end-to-end deep learning architecture
for electronic health record-based automatic mortality prediction. Scientific Reports, 8(1), 1–11.
doi:10.1038/s41598-018-24885-1.
Tseng, H., Luo, Y., Cui, S., Shi, X., & Wang, L. (2017). Hybrid deep learning framework for breast
cancer diagnosis using mammograms. Medical Image Analysis, 43, 176–185.
doi:10.1016/j.media.2017.10.004.
Zhao, Z.-Q., Zheng, P., Xu, S.-T., & Wu, X. (2019). Object detection with deep learning: A review.
IEEE Transactions on Neural Networks and Learning Systems, 30(11), 3212–3232.
doi:10.1109/TNNLS.2018.2876865.
Gu, Y., Lu, X., Cheng, J., & Wang, Y. (2017). Skin cancer detection using deep learning and time-
sequence dermatoscopic images. Proceedings of the IEEE International Conference on
Bioinformatics and Biomedicine (BIBM), 17-20. doi:10.1109/BIBM.2017.8217694.