Combo Loss: Handling Input and Output Imbalance in Multi-Organ Segmentation (1805.02798v6)

Abstract: Simultaneous segmentation of multiple organs from different medical imaging modalities is a crucial task as it can be utilized for computer-aided diagnosis, computer-assisted surgery, and therapy planning. Thanks to the recent advances in deep learning, several deep neural networks for medical image segmentation have been introduced successfully for this purpose. In this paper, we focus on learning a deep multi-organ segmentation network that labels voxels. In particular, we examine the critical choice of a loss function in order to handle the notorious imbalance problem that plagues both the input and output of a learning model. The input imbalance refers to the class-imbalance in the input training samples (i.e., small foreground objects embedded in an abundance of background voxels, as well as organs of varying sizes). The output imbalance refers to the imbalance between the false positives and false negatives of the inference model. In order to tackle both types of imbalance during training and inference, we introduce a new curriculum learning based loss function. Specifically, we leverage Dice similarity coefficient to deter model parameters from being held at bad local minima and at the same time gradually learn better model parameters by penalizing for false positives/negatives using a cross entropy term. We evaluated the proposed loss function on three datasets: whole body positron emission tomography (PET) scans with 5 target organs, magnetic resonance imaging (MRI) prostate scans, and ultrasound echocardigraphy images with a single target organ i.e., left ventricular. We show that a simple network architecture with the proposed integrative loss function can outperform state-of-the-art methods and results of the competing methods can be improved when our proposed loss is used.

• The paper introduces Combo Loss, a novel function combining Dice and weighted cross-entropy to mitigate both input and output imbalance in medical image segmentation.
• The paper shows Combo Loss substantially improves segmentation results and lowers false rates on diverse medical datasets compared to standard loss functions.
• The paper suggests Combo Loss is transferable to other medical segmentation tasks and contributes to building more automated systems.

Insightful Overview of "Combo Loss: Handling Input and Output Imbalance in Multi-Organ Segmentation"

The paper "Combo Loss: Handling Input and Output Imbalance in Multi-Organ Segmentation" addresses a pivotal challenge in medical image analysis: the simultaneous segmentation of multiple organs across different imaging modalities. This task is essential in applications such as computer-aided diagnosis and therapy planning. While deep learning techniques, notably convolutional neural networks (CNNs), have shown promise in this area, they often face two critical imbalance problems: input imbalance and output imbalance. This work introduces a composite loss function, referred to as "Combo Loss," that innovatively combines Dice and cross-entropy components to address these issues effectively.

Main Contributions

1. Combo Loss Function: The core contribution is the formulation of a new loss function that integrates a Dice-based term with a weighted cross-entropy component. The Dice term addresses the input imbalance, typical of medical segmentation tasks where class-imbalance is pronounced. The cross-entropy component, structured for binary segmentation within multi-class settings, allows for explicit control over the trade-off between false positives and false negatives. By manipulating these terms, Combo Loss enables a curriculum learning approach that gradually adapts the segmentation model to improve both training and test performance.
2. Architectural Simplification: The authors demonstrate that their proposed loss can significantly augment the performance of relatively simple deep network architectures, outperforming more complex models such as 3D U-Net and 3D V-Net. This suggests that the efficacy of segmentation may be substantially influenced by the optimization strategy rather than solely by architectural complexity.
3. Extensive Evaluation: Three diverse datasets were employed to evaluate the proposed methodology: whole-body PET scans for multi-organ segmentation, MRI prostate scans, and ultrasound echocardiography images. The evaluations encompass multiple organ types and imaging conditions, enhancing the robustness of the validation process.

Significant Findings

• Empirical Improvements: The paper reports meaningful improvements in segmentation performance using Combo Loss across the datasets tested. The authors provide comparative analyses, showing significant reductions in false positive and negative rates when Combo Loss is employed as compared to using either Dice or cross-entropy alone, or the recently introduced Focal loss. The reported Dice score improvements range from 4.6% to 8.23% for various setups, with even larger relative reductions in false rates.
• Organ-Specific Enhancements: The segmentation accuracy improvement was not confined to any single organ or dataset. For instance, in PET scans, the paper shows that despite the challenges posed by multiple missing or misclassified organs, Combo Loss enabled the segmentation model to maintain high performance even under adversarial conditions.

Implications and Future Directions

The ideas presented in this paper have multiple implications for both future research and practical applications:

• Transferability Across Modalities: The proposed loss function could be adapted for other segmentation tasks within medical imaging and potentially beyond. Its ability to balance false positives and negatives explicitly is a distinct advantage in critical health applications where these errors can have different clinical consequences.
• Towards Automated Segmentation Systems: By reducing the need for elaborate post-processing typically required to correct false segmentation results, Combo Loss contributes toward more streamlined and automated segmentation systems.
• Extension to GAN-Based Methods: Future research could examine the incorporation of Combo Loss in generative adversarial networks (GANs) that have recently gained traction in segmentation tasks for adversarial training setups.

In summary, the paper contributes significantly to the discourse on loss function design in medical image segmentation. The systematic integration of components that handle both input and output imbalances without increasing model complexity represents a pragmatic approach to improving model performance in challenging clinical settings. This work sets a foundation for subsequent exploration into dynamic loss configurations and further emphasizes the criticality of loss function design in the advancement of computer vision applications in healthcare.