Research map / Label-efficient and robust learning / Recognition and long-tail

A Unified Loss for Handling Inter-Class and Intra-Class Imbalance in Medical Image Segmentation

AAAI 2025 · Label-efficient and robust learning › Recognition and long-tail · approach: Group-robust classification · 3 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes Balance loss, combining an Inter-class Balance loss (hard-example mining on majority-class pixels scaled by minority count) and an Intra-class Balance loss (per-class easy/hard weighting, improving on TopK), used in a two-stage schedule, and evaluates it on five medical segmentation tasks with a U-Net.

Problem

Medical image segmentation suffers from inter-class imbalance (small ROIs vs. background) and intra-class imbalance (easy vs. hard samples), and existing loss functions conflate these issues and need task-specific hyperparameters that do not generalize across datasets with different imbalance degree

Approach

Attributes the inter-class problem to a soft-vs-hard FP/FN calculation bias that scales with the difference between majority and minority pixel counts, and addresses it by mining the f worst background pixels (f = foreground count) without manual class weights. Intra-class balance is handled per class by weighting easy and hard pixels separately instead of discarding easy ones; the two are linearl

Key results

• Balance loss reaches 90.7% DSC on CVC-ClinicDB vs. 87.8% for CE, and 76.9% DSC on BUSI vs. 73.5% for CE.
• On KiTS19 tumor, Balance loss achieves 66.8% DSC vs. 62.7% for the best baseline (Dice/Tversky) and 59.8% for CE.
• On the NIH pancreas dataset, Balance loss reports 77.6% DSC and 5.3 HD, the best among compared losses.

Measured effects reported in the paper

Surprising findings

• On the slightly imbalanced CVC-ClinicDB dataset, most specialized imbalance losses perform worse than plain cross entropy.
• TopK-style thresholding variants fail to train (training collapse) on some tasks depending on threshold, whereas Intra-CBL is more stable.

Limitations

• Inter-CBL is unsuitable for training from scratch, requiring a two-stage schedule with a convergence criterion.
• Evaluated only on binary 2D segmentation; multi-class extension is future work.
• Balance weight alpha and threshold t were set (t=0.9, alpha=0.5) and tuned only through ablation on the test-reported datasets.

Experimental setup

Datasets: CVC-ClinicDB BUSI KiTS19 NIH pancreas

Benchmarks: CVC-ClinicDB BUSI KiTS19 (kidney and tumor) NIH pancreas

Metrics: DSC Hausdorff Distance

Baselines: CE Focal loss Dice loss Tversky loss Combo loss Unified Focal loss RCE loss TopK loss

Builds on: Weighted Cross Entropy TopK loss Focal loss Tversky loss Unified Focal loss

Closest papers by meaning

• A Semantic Knowledge Complementarity based Decoupling Framework for Semi-supervised Class-imbalanced Medical Image Segmentation (CVPR 2025 · similarity 0.85)
• Balancing the Scales: A Theoretical and Algorithmic Framework for Learning from Imbalanced Data (ICML 2025 · similarity 0.85)
• Dist Loss: Enhancing Regression in Few-Shot Region through Distribution Distance Constraint (ICLR 2025 · similarity 0.84)
• Difficulty-aware Balancing Margin Loss for Long-tailed Recognition (AAAI 2025 · similarity 0.84)
• FairSeg: A Large-Scale Medical Image Segmentation Dataset for Fairness Learning Using Segment Anything Model with Fair Error-Bound Scaling (ICLR 2024 · similarity 0.84)
• A Tale of Two Classes: Adapting Supervised Contrastive Learning to Binary Imbalanced Datasets (CVPR 2025 · similarity 0.84)