Overview / Usage

DIATOS (Diabetic Retinopathy diagnosis model and multimodal user interface) addresses the critical issue of early detection and diagnosis of Diabetic Retinopathy (DR), a leading cause of blindness globally. The project leverages machine learning and AI to enhance diagnostic accuracy for retinal images, aiming to alleviate the burden on healthcare systems by enabling quicker, more reliable screening.

The project focuses on training a convolutional neural network (CNN), specifically a transfer learning model based on ResNet50, to classify retinal images from the Kaggle Diabetic Retinopathy Dataset. Starting with a 67% accuracy, DIATOS improves upon this by employing the Intel AI PC's Neural Processing Unit (NPU) for preprocessing and model training. The NPU accelerates computation, particularly in the preprocessing pipeline and during fine-tuning of the model, significantly boosting performance while reducing latency. ResNet50, already pretrained on the ImageNet dataset, benefits from this transfer learning approach, as it adapts to the retinal image domain more effectively.

The project implements several innovative aspects:

1. AI-Powered Diagnosis: DIATOS employs Intel's NPU to optimize the performance of the diagnosis model, making it more efficient for large-scale data and enhancing accuracy.
2. Multimodal User Interface: The interface is designed to interact with healthcare professionals, allowing easy integration into clinical workflows for diabetic retinopathy screening.
3. OpenVINO Toolkit: The Intel® Distribution of OpenVINO™ Toolkit is utilized to optimize inference, enabling the model to run effectively even on edge devices or lower-end systems.

In practice, DIATOS is poised for use in medical diagnostics environments, assisting clinicians in identifying DR at early stages. By automating the process of DR detection and offering a scalable solution for large datasets, it significantly improves screening efficiency and can be deployed in telemedicine scenarios. Its potential impact lies in early intervention, helping to mitigate vision loss in diabetic patients through timely diagnosis and treatment.

This project exemplifies the use of AI and Intel’s hardware accelerators to solve a pressing healthcare problem, contributing to more accessible, reliable medical imaging diagnostics.

Methodology / Approach

In the DIATOS project for diabetic retinopathy diagnosis, your methodology is built upon transfer learning and advanced deep learning techniques. You utilized the Intel AI PC's NPU to preprocess and accelerate model training, particularly using the ResNet50 model for diabetic retinopathy detection. Initially, your model achieved a 67% accuracy, but by fine-tuning ResNet50 on the Kaggle diabetic retinopathy dataset, you improved the performance significantly.

Here’s a detailed breakdown of your approach:

You started by preparing the Kaggle diabetic retinopathy dataset, which consists of high-resolution retinal images. These images were resized and cropped to standardize them for model input. You applied transformations like resizing, rotation, and cropping, essential steps to avoid overfitting and to help the model generalize better across unseen data.

ResNet50, a popular convolutional neural network, was used as the base model. You leveraged its pretrained weights, which were originally trained on the ImageNet dataset. Transfer learning enabled you to adapt this model to the diabetic retinopathy task by freezing the early layers (to preserve general image feature extraction capabilities) and fine-tuning the later layers for the specific task of retinal image classification.

One key advantage of your approach is using the Intel AI PC’s NPU (Neural Processing Unit) for accelerated computations. This helped speed up the model training process, making it more efficient to handle the large dataset and iterative processes involved in fine-tuning. The NPU’s capability for parallel processing enhanced the training efficiency, particularly during backpropagation and gradient computation.

You carefully evaluated the model’s performance using validation datasets, tracking metrics such as accuracy and loss via tools like TensorBoard. This allowed you to monitor the model's learning process and make necessary adjustments, such as adjusting hyperparameters or incorporating additional data augmentation techniques to avoid overfitting.

By combining the power of transfer learning, high-performance computing with Intel's NPU, and robust evaluation processes, you have created a model that significantly enhances the detection of diabetic retinopathy in medical images. Your approach reflects a modern, scalable, and efficient solution to a pressing healthcare problem.

Technologies Used

For your DIATOS project in the Intel Ambassador program, the following technologies, libraries, and tools were used:

1. Intel Technologies:

   • Intel AI PC with Neural Processing Unit (NPU): Utilized to preprocess and fine-tune the Diabetic Retinopathy diagnosis model.
   • Intel® Distribution of OpenVINO™ Toolkit: Used for inference optimization.
   • Intel® Arc™ GPU: Supported model training and performance improvements.

2. Model Architecture:

   • ResNet50: Pretrained on ImageNet and fine-tuned for Diabetic Retinopathy detection using transfer learning techniques.

3. Libraries & Frameworks:

   • PyTorch: Used for model development, training, and fine-tuning.
   • Intel Extension for PyTorch: Leveraged for hardware acceleration and optimization on Intel platforms.
   • TensorFlow: Likely used for some preprocessing or alternative model exploration.
   • scikit-learn: Potentially used for data handling and performance metrics evaluation.
   • Kaggle Diabetic Retinopathy Dataset: The primary dataset used for model training and evaluation.

4. Development Tools:

   • Conda: For managing Python environments.
   • Jupyter Notebooks: Likely used for data exploration, visualization, and experimentation with models.
   • Visual Studio Code: Common IDE for coding in Python.

5. Other:

   • Docker: Possibly for containerization and ensuring reproducible environments during development.

Documents and Presentations

Repository

https://github.com/ravisinghal033/Intel-AI-Project