Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Multimodal HIE Lesion Segmentation in Neonates: A Comparative Study of Loss Functions

Annayah Usman, Abdul Haseeb, Tahir Syed

TL;DR

This study addresses the challenge of segmenting hypoxic-ischemic encephalopathy (HIE) lesions in neonatal MRI under limited data by systematically evaluating loss functions for a 3D U-Net. It compares Dice, Dice-Focal, Tversky, Hausdorff Distance, and two proposed compound losses (Dice-Focal-HausdorffDT and Tversky-HausdorffDT) to balance overlap, class imbalance, and boundary accuracy. Compound losses generally outperform standalone losses, with Tversky-HausdorffDT achieving the best Dice and NSD and Dice-Focal-HausdorffDT minimizing MSD, demonstrating the benefit of integrating region-based and boundary-aware cues. The findings highlight the importance of task-specific loss design in small datasets and suggest that such compound losses can help approach benchmark performance with fewer training epochs. This work informs loss-function design for HIE lesion segmentation and points toward future gains via advanced architectures and zero-shot approaches in data-limited contexts.

Abstract

Segmentation of Hypoxic-Ischemic Encephalopathy (HIE) lesions in neonatal MRI is a crucial but challenging task due to diffuse multifocal lesions with varying volumes and the limited availability of annotated HIE lesion datasets. Using the BONBID-HIE dataset, we implemented a 3D U-Net with optimized preprocessing, augmentation, and training strategies to overcome data constraints. The goal of this study is to identify the optimal loss function specifically for the HIE lesion segmentation task. To this end, we evaluated various loss functions, including Dice, Dice-Focal, Tversky, Hausdorff Distance (HausdorffDT) Loss, and two proposed compound losses -- Dice-Focal-HausdorffDT and Tversky-HausdorffDT -- to enhance segmentation performance. The results show that different loss functions predict distinct segmentation masks, with compound losses outperforming standalone losses. Tversky-HausdorffDT Loss achieves the highest Dice and Normalized Surface Dice scores, while Dice-Focal-HausdorffDT Loss minimizes Mean Surface Distance. This work underscores the significance of task-specific loss function optimization, demonstrating that combining region-based and boundary-aware losses leads to more accurate HIE lesion segmentation, even with limited training data.

Multimodal HIE Lesion Segmentation in Neonates: A Comparative Study of Loss Functions

TL;DR

This study addresses the challenge of segmenting hypoxic-ischemic encephalopathy (HIE) lesions in neonatal MRI under limited data by systematically evaluating loss functions for a 3D U-Net. It compares Dice, Dice-Focal, Tversky, Hausdorff Distance, and two proposed compound losses (Dice-Focal-HausdorffDT and Tversky-HausdorffDT) to balance overlap, class imbalance, and boundary accuracy. Compound losses generally outperform standalone losses, with Tversky-HausdorffDT achieving the best Dice and NSD and Dice-Focal-HausdorffDT minimizing MSD, demonstrating the benefit of integrating region-based and boundary-aware cues. The findings highlight the importance of task-specific loss design in small datasets and suggest that such compound losses can help approach benchmark performance with fewer training epochs. This work informs loss-function design for HIE lesion segmentation and points toward future gains via advanced architectures and zero-shot approaches in data-limited contexts.

Abstract

Segmentation of Hypoxic-Ischemic Encephalopathy (HIE) lesions in neonatal MRI is a crucial but challenging task due to diffuse multifocal lesions with varying volumes and the limited availability of annotated HIE lesion datasets. Using the BONBID-HIE dataset, we implemented a 3D U-Net with optimized preprocessing, augmentation, and training strategies to overcome data constraints. The goal of this study is to identify the optimal loss function specifically for the HIE lesion segmentation task. To this end, we evaluated various loss functions, including Dice, Dice-Focal, Tversky, Hausdorff Distance (HausdorffDT) Loss, and two proposed compound losses -- Dice-Focal-HausdorffDT and Tversky-HausdorffDT -- to enhance segmentation performance. The results show that different loss functions predict distinct segmentation masks, with compound losses outperforming standalone losses. Tversky-HausdorffDT Loss achieves the highest Dice and Normalized Surface Dice scores, while Dice-Focal-HausdorffDT Loss minimizes Mean Surface Distance. This work underscores the significance of task-specific loss function optimization, demonstrating that combining region-based and boundary-aware losses leads to more accurate HIE lesion segmentation, even with limited training data.

Paper Structure

This paper contains 12 sections, 9 equations, 2 figures, 1 table.

Table of Contents

  1. Introduction
  2. Loss Functions
  3. Experiments
  4. Dataset and Benchmarking
  5. Experimental Setup
  6. Experimental Design
  7. Model Architecture and Training Protocol
  8. Evaluation Metrics
  9. Results and Discussion
  10. Conclusion
  11. Limitations
  12. Future Work

Figures (2)

  • Figure 1: Overview of the Experimental Setup. The setup processes two input modalities: 3D ADC and ZADC maps, resampling them with trilinear interpolation and normalizing their intensities separately. The concatenated 2-channel 3D input (2, 192, 192, 32) then undergoes various data augmentations. A U-Net with three encoder-decoder blocks is trained using different loss functions, and segmentation performance is evaluated using the Dice Score, MSD, and NSD Score.
  • Figure 2: Visual evaluation of predicted segmentation masks versus ground truth on Axial ADC maps across all loss functions, focusing on both extremely small and large lesion types for a holistic performance assessment.