project-proposal-2025

SkinDoctor

Abstract

According to a study, Australia is one of the countries with highest incidence rate of skin cancer in the world, with at least two-thirds of Australians diagnosed with skin cancer at some point in their lives. However, regular professional skin diagnosis at the hospital is often time-consuming and costly, making it more challenage for Australians living in non-urban areas.

Artificial intelligence technology is developing rapidly and has demonstrated amazing capabilities in the fields such as image recognition and natural language processing. Based on this technological achievement, we aim to develop a skin cancer recognition application, SkinDoctor, to quickly diagnose suspicious moles and helping users determine whether medical consultation and professional diagnosis are necessary.

Author

Name: Chi Wang

Student number: 46275262

Functionality

SkinDoctor is a web-based application where users can simply upload pictures of suspicious moles, which are then inferred by visual models to identify common skin lesions and generate risk assessment reports.

Features

• Account
  • General account information include email and password.
  • Basic information such as age, gender, skin type, and medical history are also required to offer additional context for potential verification by medical experts.
  • Multiple user levels, such as medical experts and regular users.
• Image uploading & preprocessing
  • Allow users to take images or upload images on their mobile phones
  • Automatically preprocessing uploaded image to improve the inference accuracy of visual models
• Skin lesions recognition
  • Classify common skin lesions such as melanoma, basal cell carcinoma, squamous cell carcinoma, and benign moles.
  • According to the confidence level generate risk assessment from low-risk to high-risk.
  • Offering a variety of visual AI models for users choose from to increase the accuracy of AI diagnosis, allowing weighted analysis of different models if necessary
• Visualization
  • Using visualization models such as Grad-CAM to highlight the areas of interest and mark lesion regions, explaining which features influenced the model’s decision
  • Offering medical knowledge support and display detailed information on various skin lesions, including symptoms, cases, development processes, etc.
• Historical record and tracking
  • Users can create personal profiles and save recognition records.
  • Users could track and compare the lesions of suspicious moles at different time periods for the same area.
• Experts diagnose
  • Inviting professional medical experts to participate in discussions and diagnoses.

Scope

MVP incloud:

• Account
  • General account information like email and account.
  • Basic information entry, such as age, gender, skin type, and medical history.
• Image uploading&preprocessing
  • Allow users to upload images
  • Automatic image preprocessing.
• Skin lesions recognition
  • Classify common skin lesions
  • Generate risk assessment reports.
  • Multiple visual models and weighted analysis of multiple models
• Historical record and tracking
  • Personal profiles creation, track and compare lesions of suspicious moles.

Quality Attributes

Modularity

Modularity is crucial to the implementation of the project. The whole diagnose process includes image upload, preprocessing and AI inference, etc. A modular architecture allows these functions to be separated into independent modules, making future maintenance and update more convenient. For example, each visual model offen require different data preprocessing methods and if separately develop diagnose workflow for each model would lead to redundant and complex code.

Availability

SkinDoctor requires on-demand access capability. Therefore, it must be accessible on both mobile and desktop platforms. A web-based implementation ensures compatibility across different devices, allowing users to access the platform without install additional software.

Reliability

For diagnostic reliability, all vision models should be trained on high-quality medical datasets. In cases where diagnosis is challenging for single model, multiple models should be used for joint diagnosis, with a weighted approach to generate the final report. In addition, the test accuracy of all detection models must meet a certain threshold before deploying.

Extensibility

Extensibility is also crucial for SkinDoctor to accommodate the rapid advancements in deep learning technology. More powerful visual models may emerge in the future, making it essential to have the flexibility to add or replace existing models seamlessly.

Evaluation

Modularity

In order to implement a modular architecture, any single function must be independently encapsulated. It can be verified by checking whether modules can be updated or replaced independently. In addition, each component should have strong compatibility. For example, the preprocessing module should able to integrate with any visual model without any failures.

Availability

To evaluate the availability we can checking whether the application adapts to different screen sizes and whether it provides a consistent experience for both touch interactions and mouse operations.

Reliability

For the reliability of SkinDoctor, all detection models must be trained using community-approved models and datasets. Additionally, only models with a test accuracy of over 80% will be adopted. In the future, more evaluation metrics may be introduced

Extensibility

In this project, scalability is mainly reflected in the complexity of the functional expansion. For example, intergrating new models or supporting different image formats. This means that the system needs to adopt a modular architecture, ensuring clear separation of core functionalities.

##