No abstract

Multimodal Positron Emission Tomography/Computed Tomography (PET/CT) plays a key role in the diagnosis, staging, restaging, treatment response assessment, and radiotherapy planning of malignant tumors. The complementary nature of high-resolution anatomic CT and high sensitivity/specificity molecular PET imaging provides an accurate assessment of disease status. In oncology, 18-fluorodeoxyglucose (FDG) PET/CT is the most widely used method to identify and analyze metabolically active tumors. In particular, FDG uptake allows for more accurate detection of both nodal and distant forms of metastatic disease. Accurate quantification and staging of tumors is the most important prognostic factor for predicting the survival of patients and for designing personalized patient management plans. Analyzing PET/CT quantitatively by experienced medical imaging experts/radiologists is time-consuming and error-prone. Automated quantitative analysis by deep learning algorithms to segment tumor lesions will enable accurate feature extraction, tumor staging, radiotherapy planning, and treatment response assessment. The AutoPET Challenge 2022 provided an open-source platform to develop and benchmark deep learning models for automated PET lesion segmentation by providing large open-source whole-body FDG-PET/CT data. Using the multimodal PET/CT data from 900 subjects with 1014 studies provided by the AutoPET MICCAI 2022 Challenge, we applied fivefold cross-validation on residual UNETs to automatically segment lesions. We then utilized the output from adaptive ensemble highly contributive models as the final segmentation. Our method achieved a 10th ranking with a dice score of 0.5541 in the held-out test dataset (N=150 studies).

Automated segmentation of abdominal organs plays an important role in supporting computer-assisted diagnosis, radiotherapy, biomarker extraction, surgery navigation, and treatment planning. Segmenting multiple abdominal organs using a single algorithm would improve model development efficiency and accelerate model deployment into clinical workflows. To achieve broadly generalized performance, we trained a residual UNet using 500 CT/MRI scans collected from multi-center, multi-vendor, multi-phase, multi-disease patients, each with voxel-level annotation of 15 abdominal organs. Using the model trained on multimodality (CT/MRI), we achieved an average dice of 0.8990 in the held-out test dataset with only CT scans (N=100). An average dice of 0.8948 was achieved in the held-out test dataset with both CT and MRI scans (N=120). Our results demonstrate broad generalization of the model.