Almost no one trains from scratch in 2026. Transfer learning + fine-tuning is the default workflow: download a pretrained backbone, freeze the early layers, train a small head, then unfreeze for a low-LR fine-tune pass. KerasHub gives you the catalog. This track is the recipe.
Borrowing a brain that already learned to see
Transfer learning takes a model trained on a massive dataset — ImageNet, with 14M+ labeled images across 1000 categories — and adapts what it already learned to your task. The payoff is leverage: someone else spent weeks of GPU time and millions of images teaching that network what an edge, a texture, and a shape look like. You inherit all of it for the cost of a download.
Why a feature learned on cats transfers to X-rays
A deep network learns features in a hierarchy. The early layers learn primitives — oriented edges, color blobs, simple textures — that are nearly universal across any natural image, whether it's a golden retriever or a chest X-ray. The later layers compose those primitives into task-specific concepts ("this arrangement of edges is a dog's ear"). Transfer learning keeps the universal early layers and only relearns the specialized top. That's why a backbone pretrained on photos of animals gives a genuine head start on medical imaging, satellite tiles, or crop disease — the bottom of the stack barely cares what domain you're in.
What this buys you in practice
Instead of millions of images and days of training, you reach strong accuracy with hundreds of images and minutes of training. The rest of this track is the mechanical recipe: pick a backbone (next lesson), freeze it and train a head (feature extraction), then optionally unfreeze and fine-tune. Each step is one short lesson.