Anshi Aggarwal
INNOVATOR. CREATOR. ENGINEER.
Recent advancements in ophthalmology, fueled by the integration of artificial intelligence (AI), have significantly improved diagnostic processes. One notable application is the use of AI algorithms to categorize glaucoma versus normal eye images. This development is particularly crucial in India, where the prevalence of glaucoma poses a significant public health challenge.
Glaucoma or progressive optic nerve damage is a leading cause of irreversible blindness worldwide. In India, the burden of glaucoma is particularly severe due to the country's large population and aging demographic. Efficient and accurate diagnostic approaches are essential to ensure timely intervention and effective management. In India alone, 12 million people suffer from glaucoma, with 1.5 million experiencing blindness caused by it. Shockingly, more than 75% of glaucoma cases remain undiagnosed.
To identify individuals at risk of developing glaucoma, various factors such as age, family history, ethnicity, high intraocular pressure, myopia, diabetes, cardiovascular disease, and other ocular disorders have been studied. However, traditional diagnostic methods, which involve visual field tests, optic nerve examination, and intraocular pressure measurements, are time-consuming, require specialized expertise, and often lead to delays in diagnosis, particularly in resource-limited settings. There is a growing need for a faster, more accessible, and consistent diagnostic tool.
AI algorithms have emerged as game changers in ophthalmology, revolutionizing the sorting and classification of glaucoma versus normal eye images. By analyzing extensive datasets of eye images, AI systems can learn to detect subtle changes in the optic nerve head, retinal nerve fiber layer, and other relevant structures associated with glaucoma progression. This automated analysis not only provides faster results but also ensures consistency and reliability, empowering healthcare professionals to make more informed decisions.
To address this formidable challenge, I have developed a highly sophisticated AI model that demonstrates exceptional accuracy exceeding 95% in classifying images as normal or indicative of glaucoma. This program had a reduced target class, increased data for binary classification as well as an increased training period vs current algorithms. The implementation of this model involved leveraging advanced libraries such as TensorFlow, cv2, PIL, Keras optimizer, and layers, among other supporting libraries, to construct a robust backend framework.
To train the model, a comprehensive dataset comprising 3200 images was compiled by amalgamating multiple datasets sourced from Kaggle. Subsequently, this extensive dataset was partitioned into a training set consisting of 3000 images and a testing set containing 200 images. Within the training set, the image_dataset_from_directory function provided by TensorFlow was employed to subdivide the data into a training dataset comprising 2400 images and a validation dataset consisting of 600 images.
Preprocessing techniques were then applied to standardize the image size and eliminate any blurred regions. Leveraging TensorFlow, the model was constructed using a combination of Dense layers, Conv2D, MaxPooling2D, GlobalAveragePooling2D, Flatten, Dropout, and BatchNormalization. Following the model's creation, it was trained on the 2400 images from the training dataset over a period of 15 epochs, wherein the images were presented to the model in a shuffled manner for each epoch. After training, the model's efficacy was evaluated using the 600 images from the validation dataset. The trained model was subsequently saved for future utilization in the development of the web interface.
For the front-end implementation, the Streamlit platform proved instrumental in deploying the web application. Leveraging Streamlit's compatibility with GitHub, the model, images, and requirement.txt (containing a list of all essential packages) were uploaded to the repository. A significant challenge during the development of the web app involved enabling users to upload images from their local machines and preprocessing them to ensure compatibility with the trained model. Once the final code was uploaded to GitHub, it was linked with Streamlit to meet their publication requirements, ultimately culminating in the deployment of the web application. The entire development process, from inception to completion, encompassed an estimated timeframe of 3-4 months, ensuring its readiness for adoption by ophthalmologists. The ophthalmologists can input any fundus images of the eye into the software and they will get classified into glaucoma vs normal.
The utilization of AI in sorting glaucoma vs normal eye images represents a transformative advancement in ophthalmology. In India, with its high burden of glaucoma and limited access to specialized care, AI-powered image analysis offers a revolutionary solution. By leveraging technology, healthcare professionals can achieve early detection, timely intervention, and better management of this sight-threatening condition. The integration of AI into diagnostic processes has the potential to transform the landscape of ophthalmology, empowering healthcare professionals and improving patient outcomes.
The future holds promising opportunities for further advancements in the field of glaucoma diagnosis through AI. Expanding the size and diversity of the training dataset can enhance the model's performance and generalization. Continuous refinement of image preprocessing techniques will contribute to improved data normalization and artifact removal. Additionally, incorporating more advanced deep learning architectures and exploring innovative feature extraction methods may enhance the model's ability to detect subtle changes associated with glaucoma progression.