Unsupervised Domain Adaptation for Abdominal Organ Segmentation Using Pseudo Labels and Organ Attention CycleGAN
Keywords: Unsupervised domain adaptation, abdominal organ segmentation, organ attention CycleGAN
Abstract: Abdominal organ segmentation in MRI scans poses significant challenges due to the scarcity of annotated data and the substantial domain shift between MRI and more readily available CT scans. In response to these challenges, we propose a novel approach leveraging Organ Attention CycleGAN for unsupervised domain adaptation (UDA) in abdominal organ segmentation. Our method begins by translating labeled CT images into corresponding MRI modalities using an enhanced CycleGAN model that incorporates an organ attention mechanism. This mechanism ensures the preservation of critical anatomical structures during the translation process.
Following the image translation, we employ the nnU-Net V2 framework, enhanced with Residual Encoder Presets, to perform fully supervised segmentation training on the translated MRI images. This combination allows our model to leverage the extensive labeled CT datasets effectively and adapt them to the MRI domain, achieving robust segmentation performance without requiring annotated MRI data.
To further refine the model’s performance, we introduce a self-training process using a prediction consistency algorithm. By generating multiple predictions via 5-fold cross-validation and evaluating their consistency using the Dice coefficient, we select the most reliable pseudo labels for additional training. This approach enables our model to improve segmentation accuracy on real MRI scans.
Our method was evaluated on the official validation set of the MICCAI FLARE 2024 TASK3, achieving promising results with an Organ DSC of 0.77 and an Organ NSD of 0.83, further highlighting the effectiveness of our approach in addressing the challenges of UDA for abdominal organ segmentation.
Submission Number: 18
Loading