Whole-body tumor segmentation from FDG-PET/CT: Leveraging a segmentation prior from tissue-wise projections

<b>Background</b>: Accurate tumor detection and quantification are important for optimized therapy planning and evaluation. Total tumor burden is also an appealing biomarker for clinical trials. Manual examination and annotation of oncologic PET/CT is labor-intensive and demands a high level of expertise. One significant challenge is the risk for human error, leading to potential omission of especially small tumors and tumors with low FDG uptake. <b>Purpose</b>: In this study, we introduced an automated framework with segmentation prior, from a tissue-wise multi-channel multi-angled based approach, to enhance tumor segmentation in whole-body FDG-PET/CT. <b>Method</b>: The proposed framework utilized a segmentation prior generated from tumor segmentations in tissue-wise multi-channel projections of the standardized uptake value (SUV) from PET. Projections were created from various angles and the tissues were identified based on their CT Hounsfield values. The resulting segmentation masks were subsequently backprojected into a unified 3D volume for creation of the segmentation prior. Finally, the segmentation prior was provided as an additional input channel along with the CT and SUV images to three variants of 3D segmentation networks (3D UNet, dynUNet, nnUNet) to enhance the overall tumor segmentation performance. All the methods were independently evaluated using 5-fold cross-validation on the autoPET dataset and subsequently tested on the U-CAN dataset. <b>Results</b>: Combining the segmentation prior with the original SUV and CT images improved overall tumor segmentation performance significantly compared to a baseline network. The increase in Dice coefficient for lymphoma, lung cancer, and melanoma across different segmentation networks were: 3D UNet ( <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML"> <mml:msup><mml:mrow><mml:mn>0.04</mml:mn></mml:mrow> <mml:mrow><mml:mo>⁎</mml:mo></mml:mrow> </mml:msup> </mml:math> , <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML"> <mml:msup><mml:mrow><mml:mn>0.02</mml:mn></mml:mrow> <mml:mrow><mml:mo>⁎</mml:mo></mml:mrow> </mml:msup> </mml:math> , <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML"> <mml:msup><mml:mrow><mml:mn>0.11</mml:mn></mml:mrow> <mml:mrow><mml:mo>⁎</mml:mo></mml:mrow> </mml:msup> </mml:math> ), dynUNet ( <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML"> <mml:msup><mml:mrow><mml:mn>0.05</mml:mn></mml:mrow> <mml:mrow><mml:mo>⁎</mml:mo></mml:mrow> </mml:msup> </mml:math> , <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML"> <mml:msup><mml:mrow><mml:mn>0.04</mml:mn></mml:mrow> <mml:mrow><mml:mo>⁎</mml:mo></mml:mrow> </mml:msup> </mml:math> , <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML"> <mml:msup><mml:mrow><mml:mn>0.08</mml:mn></mml:mrow> <mml:mrow><mml:mo>⁎</mml:mo></mml:mrow> </mml:msup> </mml:math> ), and nnUNet ( <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML"> <mml:msup><mml:mrow><mml:mn>0.02</mml:mn></mml:mrow> <mml:mrow><mml:mo>⁎</mml:mo></mml:mrow> </mml:msup> </mml:math> , <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML"> <mml:msup><mml:mrow><mml:mn>0.00</mml:mn></mml:mrow> <mml:mrow><mml:mi>n</mml:mi> <mml:mi>s</mml:mi></mml:mrow> </mml:msup> </mml:math> , <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML"> <mml:msup><mml:mrow><mml:mn>0.03</mml:mn></mml:mrow> <mml:mrow><mml:mo>⁎</mml:mo></mml:mrow> </mml:msup> </mml:math> ), respectively; *, p-value < 0.05; ns, non-significance. <b>Conclusion</b>: The increased segmentation accuracy could be attributed to the segmentation prior generated from tissue-wise SUV projections, revealing information from various tissues that was useful for segmentation of tumors. The results from this study highlight the potential of the proposed method as a valuable future tool for time-efficient quantification of tumor burden in oncologic FDG-PET/CT.