Professional Documents
Culture Documents
CLASSIFICATION OF
GLAUCOMA USING
DEEP LEARNING
BASED AI MODEL
Presented by:-
Ananya Singh
Aayush Srivastava
Amit Rawat
Nayanpath Jhunjhunwalla
INTRODUCTION
• Glaucoma is a leading cause of irreversible blindness. This presentation explores the use
of deep learning AI models for its identification and classification.
• Glaucoma is the second most common cause of blindness in the United States and the
single most important cause of blindness among African-Americans. Although
approximately 3 million Americans have glaucoma, only 1 million are aware that they
have the disease. About 80,000 Americans are blind from glaucoma.
• Glaucoma is a group of eye conditions that damage the optic nerve.
It can lead to vision loss if not detected and treated early.
• The changes in the visual field may manifest as visual loss either in
the nasal field, near the central field, or in the midperipheral field.
• Although elevated intraocular pressure (IOP) is a risk factor for glaucoma,
the sensitivity, specificity, and positive predictive value of IOP screening
for glaucoma is poor. Up to 50% of all patients with glaucoma may have
pressures below 22 mm Hg at any given screening; therefore, glaucoma
may remain undetected if these patients are screened on the basis of
intraocular pressure alone.
• Various openly available datasets, such as the Ophthalmology Grand Challenge (OGC)
dataset and the Singapore Eye Research Institute (SERI) dataset, can be employed for this
purpose. These datasets have been extensively utilized in glaucoma research, providing a
diverse array of retinal images and associated clinical information.\
• The Ophthalmology Grand Challenge (OGC) dataset is a well-established collection of
retinal images from both glaucoma and non-glaucoma patients. It encompasses images
captured through various imaging modalities like fundus photography and optical
coherence tomography (OCT).
DATA PREPROCESSING
• The preprocessing of the collected data is vital to enhance its quality and eliminate any
noise or artifacts. This process might involve tasks like image resizing, normalization, and
augmentation techniques to bolster the robustness and generalization of the AI model.
• One commonly used preprocessing technique is image resizing, which ensures uniformity
by standardizing the dimensions and resolution of retinal images. This aids in achieving
consistency and streamlines the training process. Additionally, it helps in reducing
computational complexity and memory requirements during model training.
• Normalization is another pivotal preprocessing step, encompassing the scaling of pixel
values of retinal images to a standardized range, typically between 0 and 1. This step
eradicates variations in brightness or contrast across different images, making them more
comparable and suitable for training the AI model.
• Augmentation techniques are employed to enhance the robustness and generalization of
the AI model by applying various transformations to the retinal images, such as rotation,
translation, scaling, and flipping. By augmenting the dataset with these transformed
images, the model becomes more resilient to variations in image orientation, position, and
size.
RESULTS AND DISCUSSION
• In this section, we present the results obtained from our deep learning-based AI model for
glaucoma identification and classification. We report key performance metrics, including
accuracy, sensitivity, specificity, and AUC-ROC.
• Additionally, the AUC-ROC metric summarizes the overall performance of the model
across all possible classification thresholds. A higher AUC-ROC value indicates superior
discrimination between glaucoma and non-glaucoma cases.
• These metrics offer a comprehensive understanding of the model's performance, covering
aspects like avoiding false positives, detecting all positive cases, and striking a balance
between precision and recall. We also compute the Receiver Operating Characteristic
(ROC).
• Moving on to cross-validation, we employ the k-fold cross-validation technique to ensure
the model's generalization performance and mitigate overfitting risks. By dividing the
dataset into k subsets, training the model on k-1 folds, and evaluating on the remaining fold.
CONCLUSION AND FUTURE SCOPE
• While the results show promise, the identification and classification of glaucoma using deep learning-
based AI models face several challenges. These encompass the necessity for extensive and varied
datasets, the model's decision interpretability, and its applicability in real-world clinical setups .
• A primary hurdle is the limited availability of comprehensive and varied datasets. Deep learning models
demand a significant volume of accurately labelled data for effective learning and generalization.
• Another obstacle lies in understanding the rationale behind the model's decisions. Deep learning models
often operate as enigmatic entities, making it challenging to comprehend the logic guiding their
predictions
• Generalizing the model's performance to real-world clinical scenarios is yet another challenge. Deep
learning models trained on specific datasets might not exhibit the same level of performance when
applied to different populations or imaging devices