What is Semantic segmentation?

Semantic segmentation is a computer-vision task that assigns a class label to every pixel in an image—road, skin, product, sky, etc so the model understands exact shapes and boundaries rather than just rough regions or boxes.

Unlike object detection, which draws rectangles, segmentation traces the true outline; unlike instance segmentation, it doesn’t separate identical objects from one another, it colors them with the same class.

Pixel-accurate understanding unlocks real work: clinicians can label tissue and tumor boundaries on medical images using DICOM annotation to measure change precisely.

Autonomous systems can read drivable space and lane edges using camera and LiDAR annotation, retailers can compute shelf share, and factories can spot hairline defects.

This precision improves safety, compliance, and downstream automation because decisions are made on true object extent, not approximations.

Example

A radiology group segments liver lesions on DICOM scans to track treatment response. After aligning on boundary rules, they label a pilot set, train a model to prefill masks, and review only uncertain slices.

With cleaner, pixel-level masks, volumetric measurements stabilize across readers, the model improves each cycle, and clinicians get reliable week-over-week change curves that stand up to audit.

‍