Hybrid organ-based loss function for deep learning-based semantic organ segmentation

Introduction: Automated and robust segmentation of the major body organs plays a critical role in clinical practices. To address this need, deep learning frameworks have been widely employed to perform organ segmentation with minimal human intervention. The deep learning algorithms tend to rely on the overall minimization of a cost function which reflects the levels of the area, volume, and/or contour overlap between the predicted and ground-truth masks/contours. This framework could be suboptimal since it may not take into account realistic or genuine shape variation of the organ of interest. In this light, deep learning-based segmentation solutions may not properly model/capture the valid shape variations, leading to gross errors and/or outliers in organ segmentation due to the lack of efficient constraints in the loss functions.  The aim of this study was to purpose a hybrid loss function for the training of the deep learning networks for the task of organ segmentation, wherein both prior knowledges regarding the shape variations and conventional overlap measurement are considered.    

Methods: Prior to the development of the deep learning solution, a shape model of the target organ was developed based on the population shape variation merged together using a distance map from the boundaries. This shape prior provides a probability/certainty index to measure the level of similarity between the estimated structure and population-based shape variations. The advantage of this shape prior knowledge is that it could be developed from any dataset or source and does not depend on the training dataset. The shape-based loss function compartment was combined with classic loss functions to take into account the levels of overlap between the estimated and reference structures. To this end, a number of loss functions were examined and the Tversky cost component exhibited superior performance together with the shape-based loss function.  The proposed loss function was evaluated against the conventional loss functions (such as Dice, Tversky, Hausdorff loss functions) for the segmentation of the kidney and Hippocampus from CT and MR images, respectively.

Results: For the segmentation of the kidney from CT images the shape-based loss function achieved an accuracy of 93.9±1.5 in terms of Dice measure compared to the Dice, Tversky, Hausdorff loss functions with Dice indices of 92.8±1.6, 92.9±1.6, and 91.5±1.7 (p-values <0.05), respectively. Similarly, the proposed loss function exhibited superior results for the task of Hippocampus segmentation from MR images with Dice index of 90.8±1.3, wherein the Dice, Tversky, Hausdorff loss functions led to Dice indices of 88.4±1.5, 88.5±1.5, and 87.3±1.6 (p-values <0.05), respectively.&nbsp;

Conclusions: The performance comparison of the neural networks with and without the shape-based loss function demonstrated the superior performance of the proposed hybrid shape-based cost function. This hybrid cost function led to no outlier in the segmentation of the kidney and hippocampus. &nbsp;

• Download figure
• Open in new tab
• Download powerpoint