Improved automated tumor segmentation in whole-body 3D scans using multi-directional 2D projection-based priors

• DOI: 10.1016/j.heliyon.2024.e26414
• Publication Date: 2024-02-15T12:55:01Z
• Authors: Sambit Tarai; Elin Lundström; Therese Sjöholm; Hanna Jönsson; Alexander Korenyu...; Nouman Ahmad; Mette A. Pedersen; Daniel Molin; Gunilla Enblad; Robin Strand; Håkan Ahlström; Joel Kullberg
• Keywords: H1-99; Cancer och onkologi; Maximum intensity projection; Science (General); Deep learning; Segmentation prior; Whole-body tumor segmentation; Social sciences (General); Q1-390; 03 medical and health sciences; Medical Imaging; 0302 clinical medicine; Medicinsk bildvetenskap; Cancer and Oncology; Medical image analysis; Radiologi och bildbehandling; Backprojection; Radiology, Nuclear Medicine and Medical Imaging; Research Article

Early cancer detection, guided by whole-body imaging, is important for the overall survival and well-being of patients. While various computer-assisted systems have been developed to expedite enhance diagnostics longitudinal monitoring, detection segmentation tumors, especially from scans, remain challenging. To address this, we propose a novel end-to-end automated framework that first generates tumor probability distribution map (TPDM), incorporating prior information about characteristics (e.g. size, shape, location). Subsequently, TPDM integrated with state-of-the-art 3D network along original PET/CT or PET/MR images. This aims produce more meaningful masks compared using baseline alone. The proposed method was evaluated on three independent cohorts (autoPET, CAR-T, cHL) images containing different forms, obtained imaging modalities, acquisition parameters lesions annotated experts. evaluation demonstrated superiority our over model significant margins in terms Dice coefficient, lesion-wise sensitivity precision. Many extremely small (i.e. most difficult segment) were missed but detected without additional false positives, resulting clinically relevant assessments. On average, an improvement 0.0251 (autoPET), 0.144 (CAR-T), 0.0528 (cHL) observed. In conclusion, TPDM-based approach can be any UNET potentially accurate robust results.