Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Learning to Look Closer: A New Instance-Wise Loss for Small Cerebral Lesion Segmentation

Luc Bouteille, Alexander Jaus, Jens Kleesiek, Rainer Stiefelhagen, Lukas Heine

TL;DR

This work tackles the problem of missing small cerebral lesions in voxel-wise Dice losses by introducing CC-DiceCE, an instance-aware loss built on the CC-Metrics framework that uses lesion-level Voronoi regions to weight per-component errors. It is evaluated within the robust nnU-Net pipeline and compared against blob loss and DiceCE across five heterogeneous brain MRI datasets. CC-DiceCE consistently improves lesion-wise detection (recall and CC-Dice) with minimal or no loss to global Dice, and generally outperforms blob loss, suggesting its role as a general instance regularizer rather than a small-lesion only enhancer. The approach is simple to implement, leverages per-lesion structure, and holds practical potential to improve radiological detection while preserving overall segmentation quality, with future work extending to additional modalities and pathologies.

Abstract

Traditional loss functions in medical image segmentation, such as Dice, often under-segment small lesions because their small relative volume contributes negligibly to the overall loss. To address this, instance-wise loss functions and metrics have been proposed to evaluate segmentation quality on a per-lesion basis. We introduce CC-DiceCE, a loss function based on the CC-Metrics framework, and compare it with the existing blob loss. Both are benchmarked against a DiceCE baseline within the nnU-Net framework, which provides a robust and standardized setup. We find that CC-DiceCE loss increases detection (recall) with minimal to no degradation in segmentation performance, albeit at the cost of slightly more false positives. Furthermore, our multi-dataset study shows that CC-DiceCE generally outperforms blob loss.

Learning to Look Closer: A New Instance-Wise Loss for Small Cerebral Lesion Segmentation

TL;DR

This work tackles the problem of missing small cerebral lesions in voxel-wise Dice losses by introducing CC-DiceCE, an instance-aware loss built on the CC-Metrics framework that uses lesion-level Voronoi regions to weight per-component errors. It is evaluated within the robust nnU-Net pipeline and compared against blob loss and DiceCE across five heterogeneous brain MRI datasets. CC-DiceCE consistently improves lesion-wise detection (recall and CC-Dice) with minimal or no loss to global Dice, and generally outperforms blob loss, suggesting its role as a general instance regularizer rather than a small-lesion only enhancer. The approach is simple to implement, leverages per-lesion structure, and holds practical potential to improve radiological detection while preserving overall segmentation quality, with future work extending to additional modalities and pathologies.

Abstract

Traditional loss functions in medical image segmentation, such as Dice, often under-segment small lesions because their small relative volume contributes negligibly to the overall loss. To address this, instance-wise loss functions and metrics have been proposed to evaluate segmentation quality on a per-lesion basis. We introduce CC-DiceCE, a loss function based on the CC-Metrics framework, and compare it with the existing blob loss. Both are benchmarked against a DiceCE baseline within the nnU-Net framework, which provides a robust and standardized setup. We find that CC-DiceCE loss increases detection (recall) with minimal to no degradation in segmentation performance, albeit at the cost of slightly more false positives. Furthermore, our multi-dataset study shows that CC-DiceCE generally outperforms blob loss.

Paper Structure

This paper contains 12 sections, 3 equations, 2 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Losses
  4. Datasets
  5. Experimental Setup
  6. Training
  7. Datasets
  8. Evaluation
  9. Results
  10. Discussion
  11. Conclusion
  12. Compliance with Ethical Standards

Figures (2)

  • Figure 1: Example from the Stanford BrainMetShare dataset sbm. Left: perfect segmentation, each lesion shown in a different color. Right: imperfect prediction. Losses are listed beneath each rendering. The brain hull is schematic and not anatomically accurate.
  • Figure 2: Per-voxel gradients with respect to logits for BlobDiceCE (left) and CC-DiceCE (right). Green rectangles denote ground truth, and light blue rectangles denote prediction. Shading color encodes gradient sign and darkness encodes magnitude. Values are normalized per panel. Voronoi regions are shown on the right. Normalized gradient values are shown in text boxes.