Post by Amy Jayant Rangari

I've wanted to work on image classification for a while now, and I finally built it on a medical dataset A two-stage deep learning pipeline that classifies fetal ultrasound images by anatomical plane — built with interpretability as a core requirement The pipeline: → Stage 1: detects whether a frame shows a standard anatomical view at all → Stage 2: if it does, identifies which of 7 planes (brain sub-types, abdomen, femur, thorax, cervix) → Every prediction comes with a Grad-CAM heatmap, so you can actually see where the model is looking Some of what I learned along the way: • Patient-level data splitting matters enormously in medical imaging (no leakage across train/val/test) • A two-stage cascade made errors far more diagnosable than a single 8-class model would have • Grad-CAM helped me catch a real concern early — whether the model was relying on a "centering shortcut" Tech stack: PyTorch + torchvision (ResNet50) for the models, Grad-CAM for interpretability, Flask for the backend, and a custom HTML/CSS/JS frontend — trained on Kaggle (GPU) and deployed on Hugging Face Spaces (Docker). 🔗 Live demo: https://lnkd.in/dGu_vx9T 💻 Code: https://lnkd.in/dbJtKpYQ 📊 Dataset: https://lnkd.in/d3NbU4h7 Would love to hear thoughts, feedback and questions on it! #MachineLearning #DeepLearning #MedicalImaging #ComputerVision #PyTorch

Video Content