Vision Spectra (Medical Diagnosis AI)

Website Demo

Inspiration

Around the world, healthcare systems are under pressure. The demand for radiology services is rising faster than the supply of trained radiologists, and in many countries, patients are left waiting far too long for critical diagnoses. We asked ourselves: Could AI help ease this burden and bring faster, more accessible care to those who need it most?

In many low- and middle-income countries, the shortage of radiologists has become a critical challenge. In India and China, the number of patients far exceeds the radiology workforce, causing serious delays in diagnosis and treatment. In parts of Africa, the situation is even more severe, some rural regions lack proper X-ray machines altogether, making it nearly impossible to detect diseases such as tuberculosis, pneumonia, or bone fractures in time.

These delays have devastating consequences. Tuberculosis remains one of the world’s top infectious killers, pneumonia is still a leading cause of childhood deaths in low-resource regions, and untreated fractures often lead to lifelong disability. Even in wealthy countries like the United States and the United Kingdom, rising demand and radiologist shortages are beginning to strain healthcare systems.

This inspired us to build an AI-powered assistant, a tool that can help radiologists diagnose quickly and accurately. By reducing the time spent analyzing X-rays, radiologists can treat more patients, prioritize urgent cases, and deliver better care, especially in places where every second counts.

What it does

Our project is designed to assist radiologists by analyzing chest X-rays automatically. The workflow is simple:

Upload a chest X-ray image through the system.
The AI model processes the image in seconds.
It returns the most likely diagnosis, along with confidence scores for each category.

Currently, our AI can detect:

Pneumonia
Tuberculosis (TB)
Pleural Effusion
Lung Nodules / Masses (potential lung cancer indicators)
Normal (healthy lungs)

By providing quick and reliable results, our tool helps radiologists save time, focus on urgent cases, and deliver better care, especially in hospitals facing radiologist shortages.

How we built it

AI Model: EfficientNetB3 (pre-trained on ImageNet)
Backend: Flask API serving PyTorch model
Frontend: Simple HTML/CSS/JS interface
Training Data: Combined 3 Kaggle datasets into one big dataset

Challenges we ran into

Different X-ray machines: Images vary by quality and resolution.
Domain shift: Models may not generalize well to data from new hospitals.
Limited data: Public datasets mostly come from big, modern hospitals.
Adoption barrier: Film-based X-rays in some regions need digitization first.
Class imbalance: Some diseases had fewer images than others.
Multi-disease training: Harder to make the model accurate across all classes.

Accomplishments that we're proud of

Finished our demo within just 24 hours, even though it was our first hackathon.
Organized the project effectively and collaborated as a team under time pressure.
Trained and deployed our own model successfully within the limited time.

What we learned

Learned how to combine multiple medical datasets into a unified training set.
Successfully trained deep learning models on medical images despite a limited dataset and time constraints.
Built a full-stack ML application in under 24 hours, connecting backend AI with a frontend interface.
Gained experience working with medical imaging data formats and adapting them for real-world use.

What's next for Vision Spectra (Medical Diagnosis AI)

AI Agent integration: Link directly with hospital medical imaging systems to receive X-rays instantly and reduce diagnosis time.
Treatment & severity insights: Extend predictions to include possible treatments or severity assessment.
Radiologist feedback loop: Collect feedback during use to continuously retrain and adapt the model for each hospital.
Dockerization: Package the system for easy deployment across different environments.