Abstract
Consistent segmentation of CT scans in COVID-19 patients across multiple time points is important to accurately evaluate disease progression and therapeutic response. In medical domains, previous interactive segmentation studies have been mainly conducted on data from a single time point. However, the valuable segmentation information from previous time points is often underutilized in assisting the segmentation of a patient’s follow-up scans. Moreover, fully automatic segmentation techniques frequently produce results that would need further refinement for clinical applicability. In this study, we propose a novel single-network model for interactive segmentation that fully leverages all available past information to refine the segmentation of follow-up scans. In the first segmentation round, our model takes concatenated slices of 3D volumes from two-time points (target and reference), employing the segmentation results from the reference time point as a guide for segmenting the target scan. Subsequent refinement rounds incorporate user feedback in the form of scribbles that rectify the segmentation, in addition to incorporating the previous segmentation results of the target scan. This iterative process ensures the preservation of segmentation information from prior refinement rounds. Experimental results obtained from our in-house multiclass longitudinal COVID-19 dataset demonstrate the effectiveness of the proposed method compared to its static counterpart, thus providing valuable assistance in localizing COVID-19 infections in patients’ follow-up scans.
