Uh, from Stevie Raymond Global University. Issues uh before that I have completed my B.Tech. And intake. In 2000, 2002. 10 and 2020 respective.

Thank you for sharing that, Dr.. Can you elaborate on what motivated your shift from your B.Tech. to pursuing a PhD, and how that transition shaped your research interests?

I have completed my.

Thanks, Dr.. Could you describe a specific research project or academic initiative you led during your PhD at Stevie Raymond Global University?

Application of. Artificial intelligence and machine learning.

To detect glaucoma, artificial.

Understood. In your research on the application of artificial intelligence and machine learning, can you walk me through a particular experiment or implementation you personally designed and what insights you gained from the results?

Thank you. When applying AI and machine learning to detect glaucoma, what dataset did you work with, and how did you ensure its quality and relevance for your research?

Is Parliament accommodation? Is a deep learning model? And CNX.

Thank you for clarifying. Since you mentioned using deep learning models like CNNs for glaucoma detection, can you describe how you selected and preprocessed the input features for your model?

Then we are. Applying the.

Okay. Once you had your image data preprocessed for the CNN, how did you evaluate the model’s performance—did you use any specific validation techniques or metrics?

Filters solutions. Activation functions to extract the features.

And we need to preprocess the data before giving. Input to the model. Send the CNN will extract the features and at the end in the dense layer it will give basic. In a flattened layer.

Thank you. After extracting features using your CNN, what was your approach for tuning the hyperparameters of your model to achieve optimal results?

Understood. After applying filters and activation functions in your CNN, what criteria did you use to determine if the features extracted were sufficient for accurate glaucoma detection?

A particular. Features which are used to detect glucoma. Come to discretion.

Thank you for explaining your feature selection. Based on your work with segmentation and feature extraction for glaucoma detection, how did you validate that your chosen features actually contributed to better diagnostic accuracy?

Actually by selecting the features.

We have.

Applied.

Thanks for clarifying. When you used segmentation techniques like U-Net, what challenges did you face in ensuring the segmented regions matched clinical expectations for glaucoma diagnosis?

Got it. What metric did you use to measure the improvement in diagnostic accuracy after selecting those features?

Thank you. When you applied feature selection, did you use any quantitative metric—like AUC, F1-score, or accuracy—to assess the improvement after selecting features?

The classification models like Support Vector machine.

Thanks for clarifying. When you applied Support Vector Machines after feature selection, what steps did you take to prevent overfitting during your model evaluation?

Providing. Glucoma detection. Achilles.

Thanks for explaining your approach. When comparing your CNN-based glaucoma detection model to baseline models, what specific baseline did you use, and how did your model’s results differ from it?

I have taken some baseline models like uh. Machine learning models. Normal classifiers. Not to use this any deep learning models. So normal classifiers we have taken as a baseline models. Like, uh. K nearest neighbor. An heavy base classifier. A random forest decision tree. Support Vector machines all classifiers we have. Pair and we get. The accuracy and calculated in the form of the confusion matrix we calculated. And the receiver operate. Think characteristics we have. Note it.

Understood. When you compared the ROC curves of your CNN model and the baseline classifiers like KNN and Random Forest, what did you observe about the true positive rates at lower false positive rates?

9 CNN.

True positive rates are more than.

Thank you. When using your CNN for glaucoma detection, did you encounter any issues with class imbalance in your dataset, and if so, how did you address them during training?

Thank you for clarifying. When you evaluated the confusion matrices for your CNN and the baseline models, were there any particular types of errors—such as false negatives—that your CNN reduced most significantly?

The other random forest or any other models? But false must debride is somehow. Uh, similar in both cases.

But we need to improve the model so that the true positive equity. Increase and. Pause negative will be reduced.

Right, so when you found that your CNN gave higher true positive rates, did you analyze any specific cases where the baseline models were failing but your CNN succeeded?

Understood. Given your goal to further increase true positive rates and reduce false negatives in glaucoma detection, what specific strategies or model adjustments would you consider implementing next to achieve this improvement?

In the model tuning.

Okay. When tuning your CNN model, did you experiment with changing the architecture itself—such as adjusting the number of layers or filter sizes—or did you focus mainly on hyperparameters like learning rate and batch size?

Like in the back propagation. We need to update the webs. Of, uh, the.

Real networks. And the learning rate we need to check. So that our model can be achieved. More accurate.

Right. When updating the weights during backpropagation, did you try different optimization algorithms—like Adam or RMSprop—and notice any impact on your model's sensitivity or convergence?

Got it. When you adjusted the learning rate during model tuning, how did you determine the optimal value for your dataset, and what impact did this have on your model's performance metrics?

And we are using generative AI like. Gap generative adversary network.

You mentioned using Generative Adversarial Networks to improve your model—can you walk me through how you integrated GANs into your glaucoma detection pipeline and what specific advantage they provided over your original approach?

Or generating synthetic images. So that those images can be trained. Quantum models.

Understood. When you generated synthetic images with GANs for training, how did you evaluate the quality and diversity of those images to ensure they contributed positively to your model's generalization?

As we know, the biomedical images are. Very lessly available. Now at day-to-day. So we need more data to. Explain the model. If the deep learning models say. Will be trained on huge number of data, so the accuracy will be increased. So in order to. Fulfill the lack. The data we need to generate some synthetic images. 10 deaths images will be similar to the original images. The generated image is similar to the original image and we need to. Train our audience if our model is trained properly. Then they finish. You will be. Increased.

Thanks for clarifying. When you incorporated those GAN-generated images into your training set, did you notice any specific changes in overfitting or generalization performance compared to training only on the original biomedical images?

And whenever we're training with synthetic images. Somewhere, a little bit, some features are not. Uh, transferred from one layer to another layer in. Need to improve the transfer of. All the important features from one layer to another layer so that we'll get the output flame.

I see. Given that some important features aren't being effectively transferred between layers when using synthetic images, what specific changes to your model or training process would you try to improve feature retention across layers?

That attention mechanism. And improve.

Understood. When you applied the attention mechanism to your model, what changes did you observe in the feature maps or output quality during validation?

Of Ganpat. And we have a plan, a cycle can model where 2 generators in. So if the majors are there. That that image should. Contain all the important features and while training to our deep learning models for classification. The accuracy should be.

It sounds like you’ve explored both attention mechanisms and cycle-consistent GANs to preserve features in synthetic images—can you describe a specific experiment where you compared classification accuracy using images from your cycleGAN versus your earlier GAN approach?

In cycle Gan. The features. Of that images. Contains all the required info. Patience so that. The enquiries can be. What in the? Traditional can is only generating the. Images. But we don't know that all features are there or not. So we performed.

A comparison.

Thanks for explaining your approach. When you evaluated the classification performance using images generated from cycleGAN versus traditional GAN, what quantitative metric did you use to compare their effectiveness, and what were the results?

In both the models. By calculating the confusion matrix after classification mode.

Thank you for sharing. In your comparison experiment, how did you validate that the images produced by the cycleGAN contained clinically relevant features needed for accurate glaucoma classification?

Understood. When you analyzed the confusion matrices from both models, what specific differences did you observe in the types of misclassifications, especially for clinically important glaucoma cases?

In clinically important glucoma cases.

You mentioned differences in feature retention for clinically important glaucoma cases—did you notice any particular patterns in false negatives or false positives when using cycleGAN-generated images versus traditional GAN images?

Images, Norman. Otherwise it is. Acneochromaticis.

Thanks for explaining the clinical significance. When measuring the cup-to-disc ratio on synthetic images, did you notice any systematic measurement errors or inconsistencies compared to real images?

In synthetic images.

I see. When you compared the cup-to-disc ratio measurements from synthetic versus real images, how did these differences impact the final classification accuracy for glaucoma detection?

Some images. I'm not having glaucoma but. They are misclassified SP. Because. The boundary of the. A pick up and up click. This are not properly defined. In the generated images as compared to the clinically defined image. So here I found some challenges. It how to improve? The cycle can model so that. Whatever the images, it will be generated. All images having the specific. Area of. A pick up and up this disk so that we can calculate the cup to disk ratio probably.

Understood, so the challenge is generating synthetic images where the optic cup and disc boundaries are realistic enough to support reliable ratio calculations. What specific modifications or data strategies have you considered to help the cycleGAN better capture these clinically important boundaries?

I think that I have a plan.

Understood. Since you have a plan to address the boundary definition issue in synthetic images, could you outline the first concrete technical step you would take to improve the cycleGAN’s ability to generate clinically accurate optic cup and disc boundaries?

Where I can implement a?

Multiple cycle gas. Where?

Alright. What is the first technical step you would take to make the cycleGAN generate images with more clinically accurate optic cup and disc boundaries?

Multiple cycle gas. Where?

You mentioned exploring multiple cycleGANs—what would be your approach for coordinating these models to ensure that each one contributes to more accurate optic cup and disc boundary generation?

Multiple cycle gas. Where?

You mentioned implementing multiple cycleGANs—what criteria would you use to decide which model’s output is most clinically reliable for calculating the cup-to-disc ratio?

Up to the. Generated emails should be equals to the real.

You mentioned implementing multiple cycleGANs to address the boundary issue—what loss functions or evaluation metrics would you use to directly encourage accurate optic cup and disc delineation during training?

Interesting idea. When using multiple cycleGANs, how would you validate that the generated images are statistically indistinguishable from real images in terms of the optic cup and disc area distributions?

We need to validate each and every generating images.

Right, so for each generated image, what specific quantitative checks or automated tools would you use to assess whether the optic cup and disc boundaries are clinically acceptable before using the images for classification?

Comparing to the clinically defined images. We'll give a comparison of. Both the images and we'll check that whether. The result of the clinically generating images are equals to the generated images or not.

Understood. When comparing these images, what threshold or quantitative criteria would you set to determine if a generated image is clinically acceptable for research or diagnostic purposes?

Until the generation. Should be equals to the clinically defined images. We'll continue the multiple cycles. And if required. We can add the attention. Network inside each site.

Including an attention mechanism could help focus the generation on boundary details—what specific form of attention would you integrate, and how would you measure its direct impact on the accuracy of optic cup and disc segmentation?

I think I will use. Uh, the explainable AI models where?

Using explainable AI models can help with interpretability—what methods would you use to visualize or quantify how well the model focuses on the optic boundaries during generation?

Whatever the. Generation is possible through the cycle gap.

Makes sense. What approach would you take to ensure that your explainable AI model’s attention maps or feature visualizations are actually highlighting the clinically relevant boundaries, rather than just focusing on easily learned artifacts?

Those models should be explained. Through AI? What AI is doing? We can't, uh, take it blindly. We need to check. The AI yours. Because gan is nothing but a. Generation models. So how it is generating we need to explain. Each and every images whether it is generating correct. Or not after generation of correctly, we need to. Check that each and every image. How they're classifying so that misclassification should not happen?

That’s clear, Professor. When reviewing these explainability results, what would you consider as convincing evidence that the model’s focus matches expert clinical interpretation, especially for challenging edge cases?

Here we are applying.

Understood. For these edge cases where the model’s explanation might not align with clinical judgment, how would you refine the training process or dataset to reduce such mismatches in future iterations?

In a images. The edge detector is there. Whenever in any RGB image if the colour is changes at that point. Some edges are there. So we need to apply the edge detection so that. Inside a. On this image the retinal image. The optic curve and optic disc boundaries should be clear with edges if the edges are clear. So that we can easily calculate the area of both. Uptick up and update this and then. Uh, cup to disk ratio will be.

You’ve highlighted the importance of edge clarity for accurate measurement—what specific edge detection technique would you prioritize for these retinal images, and how would you assess its effectiveness in delineating the cup and disc boundaries?

The cardinal age detection. I can apply already, I have tested it uh in other images. Then I have planned like that. The kernel edge detection will be. Night on the. Uptick and uptick. This uh. Edge detection models.

Okay. When you tested kernel edge detection on other images, what specific improvements or challenges did you notice in identifying the optic cup and disc boundaries compared to standard methods?

Improvements are there. As compared to other age detection models. This detection model is. More improvement, but the challenges is. Uh, like that?

Noted. What steps did you take to address the specific challenges you encountered when applying kernel edge detection to optic cup and disc segmentation?

In this.

Understood. When calculating the cup-to-disc ratio based on these detected edges, how do you ensure that small segmentation errors at the boundary don’t significantly impact your final quantitative measurements?

We need to. Apply. Different kinds of. Kernel functions. If we apply different kernel functions, it's will be a. He didn't try and get her. And at which? Function it will give accurate result. We will fix that, yeah.

Understood. Once you determine the optimal kernel function for edge detection, how do you validate that its performance remains consistent across a diverse set of retinal images with varying quality and pathologies?

That might be a challenge for. On the researchers who are working on this. Right, we need to check that.

At which point all the images.

Right, consistency across varied datasets is crucial. What metrics or validation strategies would you use to objectively compare the edge detection results produced by different kernel functions on these retinal images?

That’s a good point. When testing across all images, how do you handle cases where certain pathologies or artifacts cause the chosen kernel function to fail or produce false boundaries?

Yes, artifacts are there. For that we have used the artifact removal. On noise removal techniques.

Got it. When you applied noise and artifact removal, what specific technique or algorithm did you find most effective for cleaning up retinal images before edge detection?

We have worked on. The reflection removal techniques.

So that.

Thanks for clarifying. After applying reflection removal, did you notice any impact on the accuracy of your edge detection for the optic cup and disc—either positive or negative?

Thanks for clarifying. After implementing reflection removal, what adjustments—if any—did you need to make to your edge detection pipeline to ensure it still performed accurately on the cleaned images?

The reflection of images should be. Remote and the artifact can be removed. And the edges should be clear.

Right. Given the adjustments made for artifact and reflection removal, what steps did you take to ensure these preprocessing techniques didn’t accidentally erase subtle pathological features needed for accurate diagnosis?

Actually the preprocessing.

Steps. What we're taking.

Thanks for explaining. When you applied your preprocessing steps, did you evaluate their impact on any specific pathological cases where edge features might be faint or ambiguous?

Uh, like, uh.

Thanks for outlining your preprocessing approach. What method did you use to quantitatively evaluate that your preprocessing steps preserved clinically relevant features in retinal images—did you compare before-and-after metrics, or use expert annotation as a reference?

Changing the color model like RGB to grey.

Thank you for describing your preprocessing steps. When you applied these techniques, did you conduct any side-by-side visual inspections or quantitative tests to confirm that small but important diagnostic features were preserved in the processed retinal images?

Switching from RGB to grayscale is a common step. When you made this change, how did you assess whether any important color-based features—such as subtle vessel differences—were lost in the process?

I did not mention that I need to. Convert adds me to grayscale, but we have tested. Get Weather. The RGB model. Of the image are grayscale images. Which? Model is giving. Best accuracy? Uh, whenever we tested, we found that the RGB models having some. Color reflections. But in grayscale, the color reflection will be removed by. Reprocessing techniques. Uh, like reflection removal techniques, we can remove the reflections from the images so that that defects reduced. But in RGB image we need to apply the same reflection model whether it is what? But not I don't know. But I will try. It is my future action.

Understood. When you compare the edge detection results between grayscale and RGB images after reflection removal, what specific metrics or observations led you to conclude that grayscale offered better performance for your application?

Previously for machine learning models. Grayscale images performed better. But nowadays for deep learning, generative AI models. I don't think that we can. We need to use the grayscale models because. We can use the RGB images. But we need to improve our model. So that it can take any kind of image. That it may be AJV or. Grayscale it's better to take the RGB image because why we. Add another converter to convert the RGB to glaze scale directly. You can take the RGB images but here. The main thing is that we need to. Move the model so that it can take. Now or extract the features present in the RGB images.

That’s a thoughtful observation. When adapting your model to process both RGB and grayscale images, what changes—if any—do you anticipate needing in your feature extraction pipeline to maintain diagnostic accuracy?

Uh, every stages from starting from. The input data set, next pre processing, then image segmentation, feature extraction, classification. Among all the steps, the feature extraction is the main important. Step in feature extraction. If we extracted the features accurately, then those features will be. Uh, going to the next step of classification. So based on those features, only the classifier will classify the images as it is a normal. Or abnormal image. So here. In the feature extraction. We need to improve. Room Our.

Right, the importance of accurate feature extraction is clear. Can you give a specific example of a feature that, if not extracted properly from retinal images, would most likely lead your classifier to mislabel a pathological case as normal?

CNN models. Uh, normally the deep learning models are called as black box. So inside the black box, we need to improve. The weights, the landing rates, overall it is called as fine. Tuning so we need. To fine tune our model so that. The features should be extracted accurately and. Useful. For detection of any kind of this.

Fine-tuning your CNN models is crucial. Can you describe a situation where improper feature extraction led your classifier to mislabel a retinal image, and how you diagnosed and corrected that issue?

Uh, actually I tested on a data set which is already tested by a clinic. Are clinical doctors. So the same images without labeling I have given to. Classifier. But it classifies. And they? True positive rate is reduced but. The false positive rate improves. So.

Got it. Since you observed a reduced true positive rate but higher false positives, what specific adjustments did you make in your feature extraction or model tuning to try to address this imbalance?

That. Is good. But it is not correctly classifying images.

I see. When you noticed the classifier was not correctly labeling images, did you review any specific misclassified cases with clinical experts to identify key features your model was missing?

Yes.

Thanks for clarifying. After consulting with clinical experts about those misclassified cases, what specific changes did you make to your model or preprocessing pipeline to improve its diagnostic accuracy?

And the misclassified images. I am. Any other models? Like, uh. Our deep learning model. Inception Net. Mobile net. Those models. Team learning models. I have tested those miss message. Patiently images in this. 2 deep learning models and I found that the accuracy is improved.

When you observed improved accuracy with InceptionNet and MobileNet on those misclassified images, what differences in their architectures or training approaches do you think contributed most to the better results?

Uh, over there. The deep learning. Are more accurate with. So it would be that, uh, features are extracted. Automatically.

That’s a useful comparison. When you switched to deep learning models like InceptionNet and MobileNet, did you notice any specific types of retinal features or pathologies that were newly detected or better handled compared to your traditional machine learning approach?

We don't need to. Play any. Extra feature extraction part only. The fine tuning is. It's quiet updation. A learning rate, all those things would be important inside the team.

Right. Since you mentioned that fine-tuning is the main manual intervention needed with deep learning models, can you give a specific example of a parameter you adjusted—like learning rate or batch size—and how that change impacted your model’s diagnostic performance?

Ah, batch wise we can give because.

Uh, before applying the CNN, we're applying the kernels.

Okay. When you adjusted the batch size during training, what changes did you observe in your model’s convergence speed or its ability to generalize to new retinal images?

Means the filters. So that. In one image the filter is applied and it will found that where the exact features are there.

Got it. When you applied different kernels before the CNN, which type of kernel or filter had the biggest impact on your model’s ability to detect subtle retinal vessel abnormalities?

Right. When you selected filters for your CNN, how did you determine which filters were most effective at highlighting the critical features needed for accurate retinal abnormality detection?

And we applied. Pulling Max, pulling Everest, pulling out there. So through that. We have extracted the. Features and those features can be used for. Message.

Understood. When you used max pooling versus average pooling in your CNN pipeline for retinal images, which approach provided clearer localization of vessel abnormalities, and how did you evaluate that?

Actually, no, we found that. On the maximum is more important.

And Max cooling is. Uh, giving much more accuracy as compared to average bullion. It is also less than taking.

Thanks for clarifying. When you evaluated the results after using max pooling, did you use any quantitative metrics or visualizations to confirm that vessel features were being accurately preserved in your output?

For the same number of images I have applied, 2 * 1 is Max cooling, another is Everest cooling and I found that Max cooling giving. Morning and.

That’s clear. For your max pooling approach, did you confirm its effectiveness by comparing the classification accuracy or by visually inspecting feature maps, and what did you notice?

Thanks for explaining your findings. When you determined that max pooling improved accuracy, did you also see a reduction in false positives or false negatives in your retinal image classification results?

Calculation. Of F1 score Roc. All these things. And misclassification reduced.

Thanks for sharing your approach. When you calculated the F1 score and ROC for your max pooling model, what was the most significant improvement you observed compared to your previous configuration?

Uh, Max billing model. Provides good.

Understood. When you saw this improvement with max pooling, did you notice any trade-offs in terms of the model's ability to capture finer vessel details that might be important for borderline cases?

No, the final models, the vessel lines can be extracted. Are you using any other models?

Thanks for clarifying. Earlier, you mentioned testing Inception Net and MobileNet on misclassified images—can you describe a situation where one of these models succeeded where the other failed, and what you learned from that comparison?

But I did not find any challenges.

Thank you for clarifying. When you compared Inception Net and MobileNet on your retinal image dataset, what was the key factor that determined which model performed better for your specific classification task?

Different. Mom's birthday brand. False negative. Two negative rate. These are combined for the. And present matrix. So through the confusion metrics I have calculated. Accuracy. Sensitivity. Reference score. All these things I have compared. And the model also compared with some baseline models like. Other researchers already done with. That so I have compared. And I found that our model provides. Want to receive? As compared.

Okay. When you compared your model’s results to those baseline models from other researchers, what aspect of your architecture or training process do you believe contributed most to the improved sensitivity and accuracy?

None of. The researchers.

Thanks for explaining. When comparing your results to prior work, how did you ensure your evaluation metrics and dataset splits were directly comparable to those used by other researchers?

See in that direction. Which direction we. Saying that we need to remove the reflections from the images. In pre pressing step. If the preprocessing step is done correctly. Then.

I see, so preprocessing was a key differentiator in your approach. What specific method or algorithm did you use to remove reflections during preprocessing, and how did you validate its effectiveness on your dataset?

Should be artifact free. In that direction we have.

Right, ensuring artifact-free input images is crucial. Can you describe the specific preprocessing technique you used to remove image reflections, and how you confirmed that this step improved the downstream feature extraction accuracy?

Yes, I have used low rank not.

Understood, you used a low-rank approach for artifact removal. What criteria did you use to determine when an image was sufficiently cleaned to proceed to feature extraction?

Any metrics we know images are in the forms. Matrix. So any metrics can be. The combination of a low rank. It's apps. The no rank is the actual. Required features and the sparse is the artifact, so based on the. Low rank. I have. Even one input emails. And the info image is divided into. Two parts. One is downloading. One another in this part. So you remove the Smarts model because it can play. Activates. Only I have been with the. Low rank modern.

Any metrics we know images are in the forms. Matrix. So any metrics can be. The combination of a low rank. It's apps. The no rank is the actual. Required features and the sparse is the artifact, so based on the. Low rank. I have. Even one input emails. And the info image is divided into. Two parts. One is downloading. One another in this part. So you remove the Smarts model because it can play. Activates. Only I have been with the. Low rank modern.

That approach makes sense. How did you quantitatively measure the impact of your low-rank artifact removal on the model's classification performance?

Interesting approach. After separating the low-rank and sparse components, how did you validate that retaining only the low-rank part actually preserved the meaningful features needed for vessel segmentation?