🧿 RetinoScope
🧠 Powered by ConvNeXt for high-accuracy retinal predictions.
🖥️ Interpretability via 🔥 Grad-CAM, ⚡ Integrated Gradients, 🧩 Occlusion sensitivity, 🌐 Feature Maps.
🩺 Built to detect Diabetic Retinopathy and Glaucoma Disease.
🖥️ Interpretability via 🔥 Grad-CAM, ⚡ Integrated Gradients, 🧩 Occlusion sensitivity, 🌐 Feature Maps.
🩺 Built to detect Diabetic Retinopathy and Glaucoma Disease.
🖼️ Or choose a sample image
🔥 Grad-CAM (Gradient-weighted Class Activation Mapping):
This visualization highlights the most influential regions in the image for the model's decision. Grad-CAM uses the gradients flowing into the last convolutional layer to produce a heatmap. Areas in red represent regions of maximum importance, while cooler zones indicate lesser influence.
✅ Ideal for understanding what parts of the retina triggered the diagnosis.
This visualization highlights the most influential regions in the image for the model's decision. Grad-CAM uses the gradients flowing into the last convolutional layer to produce a heatmap. Areas in red represent regions of maximum importance, while cooler zones indicate lesser influence.
✅ Ideal for understanding what parts of the retina triggered the diagnosis.
⚡ Integrated Gradients:
This attribution method measures the influence of each input pixel by averaging the gradients as the image transitions from a baseline (e.g., black image) to the original input. Bright zones suggest strong positive contributions to the predicted class.
🔬 It’s especially helpful when assessing pixel-level sensitivity and gradual contribution patterns across retinal features.
This attribution method measures the influence of each input pixel by averaging the gradients as the image transitions from a baseline (e.g., black image) to the original input. Bright zones suggest strong positive contributions to the predicted class.
🔬 It’s especially helpful when assessing pixel-level sensitivity and gradual contribution patterns across retinal features.
🧩 Occlusion Sensitivity:
This technique systematically occludes (blocks) parts of the input image and observes the change in prediction probability. Areas that, when blocked, cause a drop in confidence are shown as hot zones.
🧠 It simulates a kind of reverse psychology—what does the model miss when certain features are hidden?
This technique systematically occludes (blocks) parts of the input image and observes the change in prediction probability. Areas that, when blocked, cause a drop in confidence are shown as hot zones.
🧠 It simulates a kind of reverse psychology—what does the model miss when certain features are hidden?
🌐 Feature Maps:
Feature maps represent the intermediate activations from the last convolutional layers of ConvNeXT. These maps visualize how the network detects patterns like edges, textures, and structures before forming a high-level understanding.
🎯 Useful for understanding how low-level retinal details are captured and built up into complex diagnostic cues.
Feature maps represent the intermediate activations from the last convolutional layers of ConvNeXT. These maps visualize how the network detects patterns like edges, textures, and structures before forming a high-level understanding.
🎯 Useful for understanding how low-level retinal details are captured and built up into complex diagnostic cues.