Inspiration
"It is God who gives real health and that is the real wealth and not pieces of gold or silver."
At Team LungHeal, we are inspired by the need to simplify and improve early detection of lung cancer using modern AI technologies. Histopathology image analysis is time-consuming and requires high expertise. We wanted to provide both patients and doctors with an easy-to-use web platform powered by AI to assist in faster, accurate diagnosis.
What the project does
LungHeal is a web application that allows doctors and patients to:
- Log in, upload histopathology lung images.
- Select from three advanced Keras-based models:
- MobileNetV2
- DenseNet121
- InceptionV3
- Get instant AI-powered diagnosis indicating:
- Benign
- Adenocarcinoma
- Squamous Cell Carcinoma
This gives users the ability to cross-check models and choose the one that provides the most confidence in their result.
How we built it
- Frontend: Built using HTML/CSS and JavaScript for a responsive and intuitive UI.
- Backend: Python with Flask to manage model loading, image preprocessing, and prediction.
- Models Used: We utilized pre-trained models (MobileNetV2, DenseNet121, InceptionV3) and converted them to
.kerasformat for compatibility and metadata richness. - Dataset: Lung and Colon Histopathological Image Dataset from Kaggle.
- Deployment: Packaged as a prototype web app demonstrating real-time image classification.
Challenges we ran into
- Managing large histopathology images efficiently.
- Ensuring model loading and predictions are fast and optimized for web use.
- UI/UX design to make model selection simple for both medical professionals and patients.
- Balancing accuracy with computation time across three models.
Accomplishments that we're proud of
- Transitioned from older
.hfile models to.keras, adding richer model metadata and easier deployment. - Successfully implemented three different deep learning models into a single web app.
- Created a clean and accessible interface that encourages use by non-technical users.
- Integrated a multi-model comparison option for better diagnostic reliability.
What we learned
- How to preprocess and handle histopathology data efficiently.
- Deployment of multiple deep learning models in a production-like web app.
- Differences in architecture performance across MobileNetV2, DenseNet121, and InceptionV3.
- The importance of UX when building healthcare-focused applications.
What's next for LungHeal
- Add explainable AI (XAI): Help users understand what features in the image led to a prediction.
- Integrate medical feedback loop: Let doctors annotate results and fine-tune models over time.
- Security & Privacy: Implement secure login and encrypted image handling for medical data compliance.
- Model Tuning & A/B Testing: Let users compare diagnostic confidence across different models.
Note for the Hackathon Team:
This project builds upon a basic model previously usingResNet.h. For this hackathon, we've added:
- Three advanced Keras models for better flexibility and accuracy.
- A new web interface tailored for patient-doctor usage.
- Improved metadata handling and selection interface to empower model choice. These features were developed entirely during the hackathon to meet the challenge requirements and push the original idea significantly forward.
Log in or sign up for Devpost to join the conversation.