A three-stage segmentation framework for lung cancer lesion isolation in three-dimensional positron emission tomography images

Abstract

Background

Positron emission tomography (PET) is a critical functional medical imaging modality for the early detection and diagnosis of cancers. PET imaging faces several challenges that hinder accurate interpretation including its inherently low spatial resolution, substantial variability in cancer lesions’ appearance, and difficulties distinguishing between the image background and benign lesions.

Methods

We propose a novel three-stage image segmentation framework to enhance the accuracy of lung cancer lesion identification and extraction from three-dimensional (3D) PET images. The first stage conducts a coarse segmentation using an encoder-decoder structure network to roughly position lesions. The second stage employs a multi-layer feature extraction network to learn the detailed characteristics of coarse segmentation results, mitigating false positives caused by localization inaccuracy. The last stage further refines the extracted features via dividing a sub-region of the lesion into foreground and background branches, reducing false positives caused by over-segmentation of edges. A novel lesion count loss function is introduced to guide the model to generate predictions during the training, ensuring that the predicted lesion counts align with the ground truth labels.

Results

The proposed method was evaluated on clinical 3D PET image datasets. Experimental results demonstrated a Dice Similarity Coefficient (DSC) of 85.35 %, Accuracy of 83.97 %, and Recall of 86.83 %. Compared to existing models applied to the same datasets, our method consistently achieved superior performance.

Conclusion

The proposed method significantly improves the segmentation performance of lung cancer lesions, implying that our method holds substantial potential for broader clinical application, even in low-resolution images.