Minimizing Hybrid Dice Loss for Highly Imbalanced 3D Neuroimage Segmentation

Area of Science:

• Medical Image Analysis
• Deep Learning
• Computational Neuroscience

Background:

• Deep learning has advanced medical image segmentation.
• Class imbalance in training data poses a significant challenge for one-stage segmentation networks.
• Existing strategies often address only one aspect of data imbalance.

Purpose of the Study:

• To investigate and compare existing strategies for handling class imbalance in medical image segmentation.
• To propose a novel hybrid loss function that jointly addresses multiple imbalance problems.
• To evaluate the efficacy of the proposed loss function in 3D brain lesion segmentation.

Main Methods:

• Comparative analysis of parallel strategies for data imbalance.
• Development of a hybrid loss function: Exponential logarithmic Dice and weighted Cross entropy Loss (EDCL).
• Implementation of EDCL with a 3D Unet architecture.
• Evaluation using 5-fold cross-validation on the Anatomical Tracings of Lesions After Stroke (ATLAS) v1.2 dataset.

Main Results:

• The proposed EDCL, integrated with 3D Unet, achieved a mean Dice score of 57.38%.
• This performance surpasses existing state-of-the-art methods on the ATLAS v1.2 dataset.
• The hybrid loss effectively handles both easy/hard example and positive/negative class imbalances.

Conclusions:

• The EDCL is an effective hybrid loss function for addressing class imbalance in medical image segmentation.
• This approach significantly enhances the performance of 3D Unet for 3D brain lesion segmentation.
• The EDCL offers a robust solution without requiring complex modifications or additional components.