(P 044) AUTOMATIC SEGMENTATION USING DEEP LEARNING TO DISTINGUISH BETWEEN 9 TYPES OF SOFT-TISSUE TUMORS WITH RADIOMICS

Objective: Soft-tissue tumors (STT) are rare tumors with a broad range of differentiation. STT subtypes differ in aggressiveness, molecular background, and optimal treatment. As differentiating STT on imaging is challenging, current clinical practice relies on invasive biopsies. To guide personalized medicine, previous studies have utilized radiomics, i.e., the use of quantitative medical imaging features, to develop noninvasive biomarkers to differentiate between various STT subtypes on MRI and CT. However, these models depend on manual 3D tumor segmentation, which is time-consuming and observer-dependent, and thus infeasible in the clinic. Here, we aim to automate STT segmentation using deep learning while retaining the radiomics model performance.

Methods: We included 520 STT patients with either a T1-weighted MRI (60 lipomas, 59 well-differentiated liposarcomas (WDLS), 72 desmoid-type fibromatoses (DTF), 61 myxofibrosarcomas (MFS), 37 myxoid liposarcomas (MLPS), and 33 leiomyosarcomas (LMS)) or CT scan (125 gastrointestinal stromal tumors (GIST), 25 leiomyomas, 25 LMS, and 23 schwannomas) from the public WORC database. Additionally, the database included 23 lymphomas and 25 gastricadenocarcinomas (GAC) with CT as these are commonly present in the GIST differential diagnosis. All tumors were manually segmented by a clinician. We used the state-of-the-art nnUNet convolutional neural network and combined several STT types to train two automatic segmentation models for respectively MRI and CT. Evaluation was performed using a 5-fold cross-validation. Performance was evaluated using the Dice similarity coefficient (DSC). Using the automatically generated segmentations, we repeated three radiomics experiments from previous studies using the WORC framework to distinguish between 1) lipoma and WDLS; 2) DTF and non-DTF; 3) and GIST and non-GIST. Patients with poor segmentation results, defined as DSC < 0.2, were excluded from the radiomics analyses as the segmentation quality was deemed insufficient. Evaluation was performed in a 100x random-split cross-validation using the area under the curve (AUC) and compared to using manual segmentations.

Results: We obtained a mean (±standard deviation) DSC of 0.71±0.34 for MRI and 0.62±0.39 for CT. Subtype-specific DSCs were 0.77±0.30 for lipoma, 0.85±0.23 for WDLS, 0.47±0.39 for DTF, 0.74±0.30 for MFS, 0.77±0.30 for MLPS, 0.75±0.33 for LMS on MRI; and 0.65±0.38 for GIST, 0.67±0.35 for leiomyoma, 0.73±0.32 for LMS, 0.70±0.39 for schwannoma, 0.50±0.46 for lymphoma, 0.31±0.33 for GAC on CT (Fig. 1). We excluded 9/114 lipoma and WDLS, 44/203 DTF and non-DTF, and 65/246 GIST and non-GIST due to DSC < 0.2. The radiomics models based on the remaining automated segmentations resulted in a mean AUC (95% confidence interval (CI)) of 0.83 (0.74, 0.91) for lipoma and WDLS, 0.75 (0.65, 0.84) for DTF and non-DTF, 0.75 (0.68, 0.82) for GIST and non-GIST (Fig. 3, Table 1). In comparison, the radiomics models using manual segmentations achieved a mean AUC (95% CI) of 0.87 (0.80, 0.94), 0.81 (0.72, 0.89), 0.78 (0.71, 0.85).

Conclusion: Our results show that deep learning can be used to generate automatic segmentation for nine different types of STT on both MRI and CT. The performance was sufficient to achieve similar but slightly lower performance in subsequent radiomics analysis compared to using manual segmentations for the distinction of lipomas from WDLS, DTF from non-DTF, and GIST from non-GIST. Hence, our automatic segmentation facilitates translation to clinical practice of these noninvasive radiomics biomarkers. As the method produced unacceptable segmentations in a part of the STTs, future work will focus on improving these segmentation methods by incorporating more STT data, including additional MRI and CT sequences, and guiding it with user interactions.