U-net与FCN语义分割对比：特征拼接对生物医学分割的作用解析

Great question—this is one of those design choices that makes U-Net such a staple in biomedical imaging, and it all ties directly to the unique pain points of segmenting medical data. Let’s break down the key reasons this feature matters so much:

• Preserves critical fine-grained spatial details
  Biomedical images live or die by small details: think cell boundaries, tiny vascular branches, or subtle tissue edges that separate healthy from diseased regions. FCNs rely on upsampling (like transposed convolution) alone to recover spatial information, which often blurs or loses these low-level, high-resolution details. U-Net’s skip connections take the cropped, high-res feature maps from the contraction path (where the model first captures edges, textures, and precise locations) and stitch them directly to the upsampled feature maps in the expansion path. This gives the model a "reference map" of exactly where those tiny structures are, instead of forcing it to guess during upsampling. For example, in cell segmentation, this is the difference between cleanly separating adjacent cells and merging them into a single blob.

• Works with limited labeled medical data
  Labeling biomedical images is incredibly time-consuming and requires expert domain knowledge (think a pathologist manually outlining every cell in a slide). U-Net’s concatenation acts as a form of feature reuse: the model doesn’t have to learn how to reconstruct spatial details from scratch during upsampling—it leverages the features it already extracted in the contraction phase. This makes the model far more data-efficient than FCNs, which need more labeled examples to learn how to recover lost spatial information. It’s one of the big reasons U-Net performs so well even on small datasets, which is the norm in medical imaging.

• Merges multi-scale features for complex medical targets
  Medical images are full of multi-scale targets: a single MRI scan might contain large organ structures and tiny lesions, or a histology slide might have big tissue regions and small individual cells. The contraction path of U-Net captures high-level semantic features (e.g., "this is a tumor region"), while the skip-connected features bring in low-level positional details (e.g., "the tumor’s exact boundary is here"). By concatenating these, the model gets the best of both worlds: it can identify what a structure is and where it is with precision. FCNs lack this direct fusion, so they often struggle to align high-level semantic predictions with precise spatial boundaries.

• Robustness to noise and low contrast in medical scans
  Many medical imaging modalities (like ultrasound, MRI, or low-light microscopy) have high noise, low contrast, or artifacts. The low-level features from the contraction path might include noise, but when concatenated with high-level semantic features, the model can filter out the noise and focus on real, meaningful edges. For example, in MRI brain segmentation, the skip connections help the model ignore motion artifacts and zero in on the actual boundaries between gray matter, white matter, and cerebrospinal fluid.

At the end of the day, this concatenation transforms FCN’s "semantic-only" prediction into a "semantic + precise localization" system—exactly what biomedical segmentation needs, where accuracy isn’t just about identifying regions, but drawing them with pixel-perfect precision.

内容的提问来源于stack exchange，提问作者Jonathan