A Novel Context Loss Function Defined on the Feature Maps: Evaluation for Lesion Segmentation from PET Images versus Conventional Loss Functions

Lesion segmentation from PET images is highly challenging due to the lack of a definite/predictable shape/morphology for the lesions compared to the anatomical organ segmentation. In this light, dedicated segmentation models should be developed for lesion segmentation from PET images. In this study, we set out to introduce a novel model and loss function in order to enhance the accuracy of lesion segmentation from PET images in comparison to conventional deep learning loss functions. To improve the accuracy of lesion delineation, context loss function was defined. In the first step, an estimate of the lesion mask is generated by the network. The predicted mask is given to a feature extractor network to generate feature maps for the predicted mask. Then, the ground truth mask is fed to the feature extractor network to generate feature maps for ground truth mask. Then an L1 loss function is defined between the two feature maps to minimize the distance between these two feature maps. This would encourage the network to detect/extract mor effective features and improve the accuracy of the lesion mask estimation.
¹⁸
F-FDG PET images of 470 patients presenting with head and neck lesions were used for training, evaluation, and testing of the proposed model. Regarding the visual inspection, the context loss function resulted in better delineation of the lesions on the PET images compared to the conventional loss functions. Considering the results of the quantitative analysis, the proposed context loss function resulted in lesion segmentation accuracy with a Dice score of 0.90 ± 0.06 compared to Dice scores of 0.86 ± 0.09 and 0.87 ± 0.05 obtained from the BCE and Tversky loss functions. Overall, the proposed context loss function led to not only improved segmentation accuracy but also no outlier was observed when this loss function was employed as opposed to the other loss functions. The proposed model could be employed in clinical practice as a decision support tool for automated lesion analysis and segmentation from PET images.