- Multi-class classification where each X-ray image is assigned to one pathology class from set of possible thoracic pathologies.
- Architecture: a custom PyTorch CNN with 5 convolution blocks, adaptive average pooling, and a small fully connected classifier head.
- Text classification pipeline built with classical machine learning, not a neural network.
- Architecture: feature engineering on comment metadata and text, TF-IDF for word/character features, plus a LightGBM-based multi-model setup for prediction.
- building a Conditional Generative Adversarial Network (cGAN) to perform a blind 4x Super-Resolution upscale. By taking the degraded 32 x 32 inputs and reconstructing the high-frequency biological textures (veins, chlorosis, and necrotic lesions) to output pristine 128 x 128 images.
The model is a custom PyTorch cGAN with a compact generator and discriminator. The generator uses a small feature extractor, residual trunk, two upsampling stages, and a sharpening head to map 32 x 32 inputs to 128 x 128 outputs. The discriminator uses a stacked CNN with adaptive pooling and a small MLP classifier. Training uses L1Loss for reconstruction, BCELoss for adversarial learning, and Adam optimizers.