(P 172) COMPARISON OF CELL SEGMENTATION METHODS IN SOFT TISSUE SARCOMA (STS) SPATIAL PROTEOMIC DATASETS

Objective: Accurate cell segmentation in spatial proteomic images is essential for analyzing the heterogeneous STS immune microenvironment. Benchmarking segmentation algorithms remains challenging due to limited high-quality, hand annotated cell masks. Recently developed reference-free quality scores enable comparison of algorithms without ground-truth annotations.

Methods: Sequential immunofluorescence was performed on tissue microarrays (1mm cores) representing 4 common STS: dedifferentiated liposarcoma [DDLPS], soft tissue leiomyosarcoma [STLMS], undifferentiated pleomorphic sarcoma [UPS], and uterine leiomyosarcoma [ULMS], using a 20-plex antibody panel on the Lunaphore Comet™ platform. Cell segmentation was performed on the images (5358 × 5358 pixels) of 27 samples based on nuclear (DAPI) and CD45 staining using four pre-trained deep learning methods: StarDist, Cellpose, MEDIAR, and CellSAM. The resulting segmentation masks and corresponding images were partitioned into non-overlapping 512 × 512 pixel tiles. Segmentation quality was assessed using a reference-free scoring framework across tiles with ≥10 cells detected by all methods. For validation, we manually annotated cell boundaries in 9 tiles (256 × 256 pixels) as the ground truth and calculated mean average precision (mAP).

Results: Among 954 image tiles evaluated, CellSAM achieved the best overall segmentation quality with a median reference-free quality score of 3.94, outperforming StarDist (3.85), MEDIAR (3.81), and Cellpose (3.85) (p < 0.001). Segmentation quality varied by sarcoma subtype with STLMS having the highest median scores (4.02), followed by ULMS (3.94), UPS (3.86) and DDLPS (3.79). In the hand-annotated tiles, CellSAM again outperformed the other methods (median mAP 0.47 vs MEDIAR 0.40, Cellpose 0.33, and StarDist 0.24; p = 0.008).

Conclusion: Reference-free quality scoring framework revealed significant differences in performance across 4 cell segmentation methods applied to STS images. CellSAM demonstrated the highest median quality score. Segmentation quality varied by STS subtypes, highlighting tissue heterogeneity’s impact on algorithm performance.