Methods for the frugal labeler: Multi-class semantic segmentation on heterogeneous labels

Area of Science:

• Biomedical image analysis
• Deep learning in medicine
• Computational biology

Background:

• Deep learning accelerates biomedical research but requires large, labeled datasets.
• Biomedical datasets are often small-scale, costly to label, and inconsistently labeled, posing challenges for supervised learning.
• Heterogeneous labels, where not all classes are labeled per sample, limit traditional supervised methods.

Purpose of the Study:

• To propose and evaluate a novel objective function for training neural networks for multi-class semantic segmentation on heterogeneously labeled biomedical data.
• To address the challenges of small-scale, sparse, and inconsistently labeled datasets in biomedical image recognition.
• To enable frugal labeling strategies in deep learning for medical imaging.

Main Methods:

• Developed a novel objective function combining class asymmetric loss and Dice loss for semantic segmentation.
• Trained neural networks on a small-scale, multi-class biomedical dataset (heartSeg) with sparse ground truth.
• Demonstrated the approach in transfer learning settings and for merging multiple heterogeneously labeled datasets.

Main Results:

• Achieved competitive results compared to standard supervised training regimes.
• Showcased the effectiveness of the proposed method on the heartSeg dataset for cardiac segmentation.
• Validated the approach's applicability in transfer learning and dataset merging scenarios.

Conclusions:

• The proposed objective function effectively trains neural networks for semantic segmentation on heterogeneously labeled biomedical data.
• Encourages the adoption of frugal labeling strategies in biomedical image recognition, reducing data requirements.
• Automating image recognition and semantic segmentation with this method supports high-throughput biomedical research.