Segmentation ability map: Interpret deep features for medical image segmentation

Sheng He, Yanfang Feng, Patricia Ellen Grant, Yangming Ou

TL;DR

Researchers have developed ProtoSeg, a parameter-free method to "peek" inside deep medical segmentation networks (like U-Net) by generating a Segmentation Ability Map (SAM) for any hidden layer. By calculating a Segmentation Ability (SA) score, they can quantify how well individual neurons or layers distinguish between an object and its background, providing a roadmap of how a network "thinks" and a way to estimate prediction quality on new images without needing ground-truth labels.

What Problem Does This Paper Address?

Deep Convolutional Neural Networks (CNNs) are the gold standard for medical image segmentation, yet they remain "black boxes." While we see the final output, we rarely understand how the hidden layers transform raw pixels into meaningful anatomical structures.

Existing interpretability tools like Attention Maps or Class Activation Maps (CAMs) are often qualitative—they show where a network looks, but not necessarily how well those internal features could actually perform the segmentation task. Furthermore, there is a lack of reliable ways to estimate the confidence of a segmentation on a test image when the "right answer" (ground truth) isn't available. This paper aims to bridge that gap by quantifying the "segmentation power" of every internal feature.

Methods

The authors introduce ProtoSeg (Prototype Segmentation), a "plug-and-play" module that requires no additional parameters.

  1. Prototype Extraction: For any given feature map in a hidden layer, the method uses the network's final output to identify which pixels likely belong to the "object" and which to the "background." It then calculates a "prototype" (an average feature vector) for both classes.
  2. Internal Segmentation: Every pixel in that hidden layer is then re-classified based on its distance to these two prototypes, creating a Segmentation Ability Map (SAM).
  3. The SA Score: By comparing this internal SAM to the actual ground truth using a Dice coefficient, the authors generate an SA Score.
  4. Offline vs. Online:
    • Offline: Used to audit a pre-trained model.
  5. Online: The SA score is added to the training loss function, forcing the hidden layers to learn more discriminative, "segmentation-ready" features earlier in the process.

Datasets

The method was validated across five diverse medical imaging tasks:

  • BraTS: Brain tumor segmentation in multi-modal MRI.
  • ISIC: Skin lesion segmentation in dermoscopic images.
  • COVID-19: Infection segmentation in lung CT scans.
  • Prostate: Whole prostate segmentation in abdominal MRI.
  • Pancreas: Pancreatic mass and parenchyma segmentation in CT scans.

Key Insights and Ideas

  • The Transition of Ability: The paper quantifies the "evolution" of a segmentation. SA scores generally increase from early layers to late layers, but there is a notable dip in "abstract layers" (the bottom of the U-Net) where spatial resolution is lowest.
  • U-Net as a Denoise Model: The authors observe that early layers are highly sensitive to noise. As the data moves through the network, the encoder path eliminates noise and the decoder path recovers the object's shape.
  • The "Oversized" Last Layer: One of the most striking findings is that the final convolutional layer is heavily redundant. Only 20% to 50% of units (neurons) actually have high SA scores. The rest are "inertia units" that contribute little to the final discriminative power.
  • Trust Without Truth: The mean SA score ($\mu$) of the final layers correlates strongly with the final Dice accuracy. This means $\mu$ can serve as a proxy for quality, allowing a system to "reject" or flag poor-quality segmentations for human review in a clinical workflow.

Key Results

  • Performance Stability: Training with the "Online ProtoSeg" loss did not hurt final accuracy; in fact, it slightly improved the Dice score across most datasets (e.g., BraTS improved from 0.8471 to 0.8514).
  • Internal Discriminative Power: On the BraTS dataset, the best individual unit in the final layer achieved an SA score of ~0.85, which is nearly identical to the performance of the entire combined output.
  • Sample Rejection: By using the mean SA score to reject the bottom 20% of "difficult" images, the average Dice score on the remaining COVID-19 test set jumped from 0.4880 to 0.5400, demonstrating its utility as a confidence metric.

Limitations and Future Work

  • Metric Specificity: The study primarily uses the Dice score as the basis for the SA score. While standard, other metrics like Jaccard index or Hausdorff distance weren't explored as internal benchmarks.
  • Dimensionality: The current implementation was evaluated on 2D network models. The authors note that extending this to 3D networks is a necessary next step for many medical applications.
  • Weight Correlation: The researchers found no direct correlation between a unit's SA score and its learned weight in the final layer, suggesting that the network doesn't necessarily "know" which of its internal features are the most accurate segmenters. Future work could involve using SA scores to prune redundant "inertia units" for more efficient models.

Summary updated

BibTeX
@article{he2023segmentation,
  author = {Sheng He and Yanfang Feng and Patricia Ellen Grant and Yangming Ou},
  title = {{Segmentation ability map: Interpret deep features for medical image segmentation}},
  journal = {Medical Image Analysis},
  year = {2023},
  doi = {10.1016/j.media.2022.102726},
}
Figure