Segment anything model (SAM) for metastatic tumor segmentation: evaluating its impact on chemotherapy response prediction

Accurate tumor segmentation is a critical step in early-stage assessment of chemotherapy effectiveness. However, it is also a technical challenge in clinical practice due to high tumor heterogeneity and ambiguous tumor boundaries. Meanwhile, foundation model-based segmentation approaches may offer a viable solution to this challenge. The purpose of this study is to evaluate the effectiveness of the segment anything models (SAM) on metastatic tumor segmentation and its impact on the performance of the following response prediction model. To achieve this, zero-shot segmentation was performed using the medical SAM (MedSAM) model on a dataset consisting of 100 advanced-stage ovarian cancer patients. Based on the segmented results, a total of 1032 radiomics features were computed, and a support vector machine-based model was established for response prediction. The segmentation performance was evaluated using a number of indices (e.g. Dice Similarity Coefficient (DSC) and the Jaccard Index (JI)). The results demonstrate that MedSAM achieved satisfactory accuracy on metastatic tumor segmentation, and it performs better on the relatedly larger tumors. When using tumors segmented by the MedSAM approach, the accuracy of the prediction model is comparable to that achieved using the same model established with the baseline segmentation tumors. This study provides valuable information for the development and optimization of SAM-based computer aided diagnosis schemes in the future.