Calibrating segmentation networks with margin-based label smoothing.

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Computer Vision

Background:

• Deep neural networks (DNNs) excel in visual recognition but suffer from poor calibration, leading to over-confident predictions.
• Standard cross-entropy loss exacerbates miscalibration by creating large pre-softmax activations for correct classes.
• Existing calibration losses, often maximizing prediction entropy, show promise in classification but are unexplored for medical image segmentation.

Purpose of the Study:

• To investigate the impact of entropy-maximizing loss functions on medical image segmentation network calibration.
• To propose a novel, flexible loss function for improved calibration and discriminative performance in medical image segmentation.
• To provide a unifying constrained-optimization perspective on existing calibration losses.

Main Methods:

• Developed a constrained-optimization framework to analyze current state-of-the-art calibration losses.
• Proposed a generalized loss function based on inequality constraints to control logit distances, introducing a controllable margin.
• Conducted comprehensive experiments on public medical image segmentation benchmarks.

Main Results:

• The proposed margin-based loss function achieved state-of-the-art results in network calibration on medical image segmentation tasks.
• The method also demonstrated improvements in discriminative performance compared to existing approaches.
• The constrained-optimization perspective revealed limitations of equality constraints in current calibration losses.

Conclusions:

• The novel inequality-constrained loss function effectively addresses the miscalibration problem in deep learning models for medical image segmentation.
• The proposed method offers a superior balance between calibration and discriminative performance.
• This work opens new avenues for developing better-calibrated deep learning models in medical image analysis.