Inspiration
The inspiration for this project stems from the critical need for speed in clinical environments. Medical professionals often face high-pressure scenarios where manual analysis becomes a bottleneck. We wanted to see if we could combine the "cold" precision of AI with the warmth of patient care to create a tool that reduces human error and speeds up the path to treatment.
What it does
Disease Classification: Using a symptom-based engine, the system categorizes potential illnesses based on reported patient data, acting as a preliminary triage tool.
Bone Age Estimation: Using computer vision, the model analyzes hand radiographs to predict a patient's skeletal age, crucial for diagnosing growth disorders.
How we built it
We took a hybrid approach to development, utilizing the strengths of the industry's two heavy hitters: PyTorch and TensorFlow.
Disease Prediction: This was treated as a multi-class classification problem using the Disease-Symptom Dataset.
Bone Age Model: We implemented a Deep Convolutional Neural Network (CNN) to perform regression. The model was trained on the RSNA Bone Age Dataset.
Challenges we ran into
One of the biggest hurdles was data heterogeneity. Medical X-rays vary significantly in lighting and orientation, requiring heavy preprocessing and augmentation in PyTorch to ensure the model didn't overfit to specific machine signatures.
Another challenge was class imbalance within the Disease-Symptom dataset. Some rare diseases had very few samples, making it difficult for the model to "learn" them without bias. We had to experiment with weighted loss functions and synthetic data oversampling to ensure the classifier remained fair and accurate across the board.
Accomplishments that we're proud of
We are incredibly proud of the inference speed we achieved. By optimizing our models, we managed to get bone age predictions down to sub-second processing times. Additionally, seeing the Disease Classifier correctly identify rare symptom clusters during testing was a massive "win" for the team.
What we learned
This project taught us that Framework Interoperability is a superpower. Using PyTorch for the rapid prototyping of the vision model and TensorFlow for the structured data pipeline allowed us to leverage the best features of both ecosystems. We also gained a deep appreciation for the ethical side of AI.
What's next for Disease Predictiona and Bone Age Estimation Model
The current iteration is just the beginning for Team Arctic. Our roadmap includes:
Expansion: Integrating more diverse datasets to include a wider range of ethnic and demographic skeletal structures.
Mobile Deployment: Converting our models to TensorFlow Lite to allow for edge computing in remote clinics with limited internet access.
Explainability: Implementing Grad-CAM (Gradient-weighted Class Activation Mapping) so doctors can see exactly which parts of an X-ray the AI is looking at when it makes a prediction.
Log in or sign up for Devpost to join the conversation.