AI Powered MediKit

• 

  User Interface, Providing lot of features

• 

  Detail Page

• 

  Classifying Data with subtle features using Vision Transformer

Inspiration

To be frank, this project has no specific inspiration; it's an outpouring of oiur thoughts. The idea was to bundle various services together into a compact and user-friendly solution.

What it does

The AI-Powered MediKit enhances medical analysis with several advanced features. Most of its functionalities are classification-based, but instead of relying on conventional CNNs, we've incorporated Vision Transformers (ViTs) for model training. While CNNs and CNN-based networks like ResNet excel in detecting simple features, they struggle with the subtle details often present in medical images. Vision Transformers, however, excel at capturing these delicate features. Additionally, the MediKit offers Heartbeat Analysis using MFCC to classify heartbeats, which helps in identifying abnormalities. We also provide few-shot classification for tablets, minimizing the need for retraining the network with new data. The project includes a herbal solution feature where users can inquire about herbal remedies.

How we built it

The development process involved analyzing the risk of misidentifying features in medical images, which can be critical. We compared the performance of ViTs with CNN-based networks and found that ViTs perform better for subtle features. Reference

Challenges we ran into

We faced challenges balancing development time with practical implementation. Additionally, a lack of data posed another significant challenge.

Accomplishments that we're proud of

The results have been impressive, with all models achieving accuracy above 93%, and some surpassing 95%.

What we learned

Tuning and tweaking are crucial for optimizing performance. ViTs perform better even for data with high intraclass variance, while CNNs excel in few-shot classification. We've learned that neither model is inherently superior; their effectiveness depends on the specific application.

What's next for AI-Powered MediKit

We plan to integrate few-shot classification across all functionalities (which may be challenging but worth pursuing) and expand the MediKit with additional medical tasks.

Built With

• python
• pytorch