Comparison of Loss Functions for Robust Deep Learning-based Echocardiography Segmentation when Learning with Partially Labelled Data from Multiple Domains

2026-07-06 · Source: Takara TLDR - Daily AI Papers · Field: Science & Research — Health & Medical Research, Mathematics & Computational Sciences · Depth: Expert, quick

Summary

A comprehensive study evaluates three loss functions—adaptive categorical cross entropy (aCCE) loss, marginal loss, and adaptive binary cross entropy (aBCE) loss—for robust deep learning-based echocardiography segmentation. The research addresses challenges in training automated cardiac structure segmentation models using partially-labelled data from multiple domains. Experiments compared these loss functions across intra-domain and inter-domain tasks, considering scenarios with single or multiple partial-labels and varying proportions of fully-labelled data. All three loss functions demonstrated strong performance in intra-domain segmentation. For inter-domain tasks with one missing label, aBCE and marginal losses showed superior results. In more complex scenarios involving multiple missing labels, marginal loss significantly outperformed the other methods, highlighting its robustness. This investigation is the first to compare loss-based solutions for multi-domain, partially-labelled echocardiography segmentation.

Key takeaway

For Machine Learning Engineers developing echocardiography segmentation models with multi-domain, partially-labelled datasets, your choice of loss function significantly impacts performance. If you encounter scenarios with multiple missing labels, prioritize marginal loss for its proven robustness. For single missing labels, aBCE or marginal loss are strong contenders. This selection optimizes model accuracy and reduces the need for extensive full data annotation.

Key insights

The marginal loss function offers superior robustness for multi-domain echocardiography segmentation with multiple missing labels.

Principles

• Loss function choice is critical for multi-domain partial-label learning.
• Different loss functions excel in specific partial-label scenarios.

Method

The study compared aCCE, marginal, and aBCE losses across intra-domain and inter-domain echocardiography segmentation tasks, varying partial-label counts and fully-labelled data proportions.

In practice

• Prioritize marginal loss for multi-label missing scenarios.
• Consider aBCE or marginal for single missing labels.
• All three losses are effective for intra-domain tasks.

Topics

• Echocardiography Segmentation
• Deep Learning
• Loss Functions
• Partially-Labelled Data
• Multi-Domain Learning
• Cardiac Imaging

Best for: Computer Vision Engineer, AI Scientist, Machine Learning Engineer, Research Scientist

Related on AIssential

• Polynomial Dice Loss for Medical Image Segmentation — Polynomial Dice Loss enhances medical image segmentation by controlling loss shape via Taylor expansion for improved performance.
• Rethinking Noise-Robust Training for Frozen Vision Foundation Models: A Cross-Dataset Benchmark with a Case Study of Small-Loss Failure — No universal noisy-label learning method exists for frozen Vision Foundation Models; selection must be regime-aware.
• 20+ Types of Loss Functions in Machine Learning — Selecting the appropriate loss function is crucial for effective model training across diverse machine learning tasks.
• EAGT: Echocardiography Augmentation for Generalisability and Transferability — Geometric data augmentations significantly improve echocardiography model generalizability across diverse datasets, outperforming intensity-based methods.
• Accuracy Improvement of Semi-Supervised Segmentation Using Supervised ClassMix and Sup-Unsup Feature Discriminator — Improved semi-supervised segmentation uses supervised ClassMix and a feature discriminator to enhance accuracy and mitigate pseudo-label inaccuracies.
• Multi-Stage Bi-Atrial Segmentation Framework from 3D Late Gadolinium-Enhanced MRI using V-Net Family Models — A multi-stage V-Net framework with MCLAHE preprocessing effectively segments bi-atrial regions from 3D LGE-MRI.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Takara TLDR - Daily AI Papers.