User-Guided Domain Adaptation for Rapid Annotation from User Interactions: A Study on Pathological Liver Segmentation

Mask-based annotation of medical images, especially for 3D data, is a\nbottleneck in developing reliable machine learning models. Using minimal-labor\nuser interactions (UIs) to guide the annotation is promising, but challenges\nremain on best harmonizing the mask prediction with the UIs. To address this,\nwe propose the user-guided domain adaptation (UGDA) framework, which uses\nprediction-based adversarial domain adaptation (PADA) to model the combined\ndistribution of UIs and mask predictions. The UIs are then used as anchors to\nguide and align the mask prediction. Importantly, UGDA can both learn from\nunlabelled data and also model the high-level semantic meaning behind different\nUIs. We test UGDA on annotating pathological livers using a clinically\ncomprehensive dataset of 927 patient studies. Using only extreme-point UIs, we\nachieve a mean (worst-case) performance of 96.1%(94.9%), compared to 93.0%\n(87.0%) for deep extreme points (DEXTR). Furthermore, we also show UGDA can\nretain this state-of-the-art performance even when only seeing a fraction of\navailable UIs, demonstrating an ability for robust and reliable UI-guided\nsegmentation with extremely minimal labor demands.\n