The following is a summary of “Automated segmentation of lesions and organs at risk on [68Ga]Ga-PSMA-11 PET/CT images using self-supervised learning with Swin UNETR,” published in the February 2024 issue of Oncology by Yazdani et al.
Prostate-specific membrane antigen (PSMA) PET/CT imaging is pivotal in quantitative image analysis, particularly within the realm of radioligand therapy (RLT) for metastatic castration-resistant prostate cancer (mCRPC). The exploration of elusive features influencing PSMA biodistribution necessitates the meticulous analysis of segmented organs at risk (OAR) and lesions. Manual segmentation, though thorough, is marred by its time-consuming and labor-intensive nature, hence the burgeoning demand for automated segmentation methods. In response, the researchers introduce a novel approach employing shifted windows UNEt TRansformers (Swin UNETR) for fully automated segmentation. Integral to this methodology is a self-supervised framework, where the model’s encoder is adeptly pre-trained on an ample repository of unlabeled data before undergoing fine-tuning encompassing the entire model, including its decoder, using meticulously annotated data.
Methodologically, the study draws upon 752 whole-body [68Ga]Ga-PSMA-11 PET/CT images sourced from two distinct centers. Leveraging 652 unlabeled images for self-supervised model pre-training, the supervised training phase capitalizes on the remaining 100 images meticulously labeled. Employing 5-fold cross-validation, wherein 64 images are allocated for model training and 16 for validation from a single center, the study group evaluates the model’s performance on a testing cohort comprising 20 hold-out images, equally distributed between the two centers. Evaluation metrics, including image segmentation and quantification, are meticulously scrutinized against ground-truth segmentation conducted by a seasoned nuclear medicine physician.
The results underscore the efficacy of their model in generating high-fidelity OARs and lesion segmentation, particularly in lesion-positive scenarios such as mCRPC.
Noteworthy is the significant enhancement in the average dice similarity coefficient (DSC) across all classes by approximately 3% owing to self-supervised pre-training, outperforming the established nnU-Net model by a substantial 5% margin. Notably, their model’s prowess in amalgamating self-supervised pre-training with supervised fine-tuning is markedly evident, particularly in the context of PET/CT inputs. The ensuing model exhibits remarkable performance with the lowest DSC for lesions at 0.68 and the highest for liver at 0.95.
In conclusion, the study heralds the development of a cutting-edge neural network meticulously primed through self-supervised pre-training on comprehensive sets of [68Ga]Ga-PSMA-11 PET/CT images, followed by fine-tuning on a judiciously curated dataset of annotated images.
The resultant model boasts commendable proficiency in generating superior-quality OARs and lesion segmentation, thus auguring well for diverse clinical applications, including the refinement of RLT and the facilitation of patient-specific internal dosimetry.
Source: cancerimagingjournal.biomedcentral.com/articles/10.1186/s40644-024-00675-x