(P 231) MINIMALLY INTERACTIVE SEGMENTATION OF SOFT-TISSUE TUMORS ON MRI AND CT USING DEEP LEARNING

Objective: Segmentations have many applications in soft-tissue tumor (STT) healthcare, as they can be used for clinical measurements, e.g., to evaluate therapy response (i.e., RECIST), to target radiotherapy, and are often required to extract quantitative imaging biomarkers (radiomics). Manual segmentation, although accurate, is time-consuming and observer-dependent, and thus not feasible in a radiologist's clinical workflow. Automatic methods, on the other hand, may have limited accuracy due to the heterogeneous nature of STTs. In this study, we present a minimally interactive segmentation method, in which the radiologist guides a deep-learning network to generate segmentations.

Methods: Our interactive method requires the user to manually mark six points near the extreme boundaries of the tumor, i.e., two extreme points for all three directions, in the image (Fig 1). The points are used to extract a region of interest (ROI) around the tumor and are transformed into an exponentialized geodesic distance map (EGD), which indicates the location of the tumor by calculating the distance in contrast between the interactions and surrounding tissue. Together the ROI and EGD serve as input for a Convolutional Neural Network (CNN). We evaluated our interactive segmentation method in the public WORC database, which included 514 patients with 11 STT phenotypes on either CT or T1-weighted MRI (Fig 2). We trained separate networks for MRI and CT and split the data in 80% for training and 20% for test. Next, we validated the interactive method in another public dataset from McGill University, referred to as the TCIA dataset, which included 51 patients with 7 STT phenotypes on CT, T1- and T2- weighted MRI. Moreover, 5 STT phenotypes included in the TCIA dataset were not presented in the 11 STT phenotypes of the WORC database. Therefore, the TCIA dataset was also used to assess the performance for STT phenotypes not seen during training of the methods. We compared the interactive segmentation method to automated segmentation using the nnU-Net framework. Performance was measured using the Dice Similarity Coefficient (DSC). The DSCs of the segmentation methods were compared with paired two-sided t-tests. Additionally, we evaluated the clinical utility of the segmentation methods by calculating correlation scores between volumetric and diameter measurements compared to ground truth segmentation.

Results: For interactive segmentation, we obtained a mean (±standard deviation) test DSC of 0.85±0.11 for CT and 0.84±0.12 for T1-weighted MRI on the WORC test dataset (Fig 3). The automatic segmentation method obtained significantly worse results (CT: 0.52 ± 0.43 (p < 0.001), T1-weightedMRI: 0.71 ± 0.35 (p=0.007)). Similarly, volume and diameter calculations based on the interactive segmentation had a better agreement (volume: r=0.96, diameter: r=0.99) than on automatic segmentation (volume: r=0.76, diameter: r=0.62) when compared to the ground truth segmentation. Validation of the interactive segmentation method on TCIA data showed similar performances in terms of DSC on all sequences (CT: 0.81 ± 0.08, T1: 0.84 ± 0.09, T2: 0.88 ± 0.08), while automatic segmentation performed worse (CT: 0.38 ± 0.341, T1: 0.81 ± 0.234, T2: 0.78 ± 0.216). Here, interactive segmentation significantly outperformed automatic segmentation on CT and T2-weighted MRI, while providing similar results on T1-weighted MRI (CT: p< 0.001, T1-weighted MRI: p=0.14, T2-weighted MRI: p=0.004).

Conclusion: Our minimally interactive model generated semi-automatic segmentations with higher DSC than automatic segmentation for various STT types on T1-weighted MRI and CT. Additionally, we showed that the interactive segmentation method can be used to accurately perform volume and diameter measurements. Furthermore, minimally interactive segmentation can be used to detect unseen tumor types and STT on modalities not seen during the training of the model. Based on these results, we are currently investigating the use of interactive segmentation in the clinical workflow.