gms | German Medical Science

Introduction: Cancer remains a predominant global health concern, with lung cancer alone responsible for 350 deaths daily in the U.S. While medical imaging (MI) and histological examinations are fundamental diagnostic tools, manual analysis proves time-intensive and costly. Deep learning (DL) offers promise for automation but faces challenges like data scarcity and domain shifts across diverse MI modalities. Typical domain shift example in MI is inter-scanner scenario, where different machine models result in voxel intensity distribution discrepancies. For inter-cancer case this dissimilarity stems from having various cancer types. This scenario presents additionally distinct spatial characteristics, such as variations in lesion size or location. Usually model trained on one domain performs worse when applied to another domain. This problem can be addressed by transfer learning methods. In this study we perform adaptation by harnessing only unlabeled data through semi-supervised learning and Unsupervised Domain Adaptation (UDA) techniques.

Methods: We present nn-Match, an innovative pipeline that merges a semi-supervised learning approach rooted in consistency regularization, inspired by MeanTeacher [1], with the state-of-the-art medical segmentation method nn-UNet [2], serving as its backbone model. The dataset employed in this study originates from the AutoPET challenge [3], comprising PET/CT scans from patients diagnosed with melanoma, lymphoma, or lung cancer. For each experimental setup, one cancer type is the source domain, while another is the target domain.

Initially, the nn-UNet model is trained on the source dataset, followed by its architecture and weights extraction. We employ consistency regularization to adapt the model to the unlabeled target domain data. This involves creating two distinct views of an input image by applying varied perturbations. These views are then processed by separate networks (student and teacher), each an instance of the previously trained nn-UNet model. Outputs from these networks are compared using a consistency loss, specifically the Dice loss. Subsequently, the student network instance undergoes backpropagation, leading to its adaptation to the target dataset. To generate labels for the target samples, we incorporate pseudo-labeling, further refined using confidence thresholds. Augmentations like additive Gaussian noise, Gaussian blur, flipping, rotation, and Cutout [3] are employed to diversify the views of each image.

Results: The nn-Match pipeline substantially enhances segmentation accuracy, as evidenced by the Dice score, across three of six inter-cancer setups. The most significant improvements occur when adapting from lung cancer to melanoma (0.377 to 0.543) and from lymphoma to melanoma, (0.577 to 0.652), where fully-supervised model scores 0.684.

Discussion: This study delves into UDA methodologies for PET/CT data, specifically addressing inter-cancer domain shifts. Our consistency-based techniques improve target domain performance without labeled data, allowing more accurate cancer tissue annotation. The results underscore the applicability of these methods to PET/CT data and paves the way for further research to validate them across conventional inter-scanner domain shift scenarios. However, challenges remain. Firstly, it relies on numerous hyperparameters, necessitating extensive experimental exploration. Secondly, the absence of target labels complicates overfitting identification, as our validation metric relies on pseudo-labels from the teacher network, leaving the optimal training termination point uncertain.

The authors declare that they have no competing interests.

The authors declare that an ethics committee vote is not required.