KD OCT ConvNeXtV2

--- license: mit pipeline_tag: image-classification tags: - medical-imaging - oct - retinal-imaging - knowledge-distillation - computer-vision --- # KD-OCT: Knowledge Distillation for OCT Image Classification This model is part of the **KD-OCT** project, which provides a framework for efficient, clinical-grade retinal Optical Coherence Tomography (OCT) image classification. It was introduced in the paper [KD-OCT: Efficient Knowledge Distillation for Clinical-Grade Retinal OCT Classification](https://huggingface.co/papers/2512.09069). The official implementation is available at the [GitHub Repository](https://github.com/erfan-nourbakhsh/KD-OCT). ## Model Description KD-OCT is a novel knowledge distillation framework designed to compress a high-performance **ConvNeXtV2-Large** teacher model into a lightweight **EfficientNet-B2** student model. The framework is optimized to classify OCT B-scans into clinical categories: - **Normal** - **Drusen** - **Choroidal Neovascularization (CNV)** - **Diabetic Macular Edema (DME)** (in the 4-class UCSD variant) By employing real-time distillation with a combined loss (soft teacher knowledge transfer and hard ground-truth supervision), the student model achieves near-teacher accuracy while significantly reducing model size and inference time, facilitating deployment on edge devices for AMD screening. ## Training Details - **Framework**: PyTorch - **Architecture**: EfficientNet-B2 (Student) distilled from ConvNeXtV2-Large (Teacher) - **Datasets**: Noor Eye Hospital (NEH) and UCSD OCT datasets - **Training Method**: Knowledge Distillation using combined KL Divergence and Cross-Entropy/Focal Loss - **Optimizer**: AdamW with Cosine Annealing ## Usage Below is an example of how to perform inference with the model. Note that according to the paper, the images should be resized to 384x384 for inference. ```python import torch from torchvision import transforms from PIL import Image # Load model model = torch.load("model.pth") model.eval() # Prepare image transform = transforms.Compose([ transforms.Resize((384, 384)), transforms.ToTensor(), transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]) ]) # Assume 'image' is a PIL image input_tensor = transform(image).unsqueeze(0) # Inference with torch.no_grad(): output = model(input_tensor) prediction = torch.argmax(output, dim=1) ``` ## Cita...