## TL;DR
Image segmentation partitions images into meaningful regions — semantic (class per pixel), instance (object per pixel), or panoptic (both). U-Net dominates medical imaging; SAM enables general-purpose interactive segmentation.

## Core Explanation
Encoder-decoder architectures: the encoder compresses spatial information into feature maps; the decoder upsamples to pixel-level predictions. Skip connections (U-Net) preserve fine spatial details lost during downsampling. Dilated/atrous convolutions maintain receptive field without resolution loss.

## Detailed Analysis
Mask R-CNN added a small FCN branch to Faster R-CNN for instance mask prediction — the dominant model from 2017-2023. SAM's three-component design: image encoder (ViT), prompt encoder, and mask decoder. The lightning-fast mask decoder enables real-time interactive segmentation after one-time image encoding.

## Further Reading
- Papers With Code: Semantic Segmentation
- Meta AI: SAM Demo & Research
- MONAI: Medical Open Network for AI