We used the openly available TotalSegmentator dataset [18], a CT image segmentation dataset. The TotalSegmentator dataset is released under the Creative Commons Attribution 4.0 International license, permitting unrestricted reuse for research. The original CT images were collected retrospectively at University Hospital Basel. The Ethics Committee Northwest and Central Switzerland (EKNZ) approved a waiver of ethical approval for that retrospective study (BASEC Req-2022–00495). No additional ethics review was required for this secondary analysis. Patient consent was waived by the EKNZ due to the deidentified, retrospective nature of the original data collection. All CT images in the TotalSegmentator dataset were fully deidentified of any protected health information before public release. The dataset contains anonymized images and no patient identifiers. No financial or other compensation was provided to participants for the original data collection, and none was provided for this secondary analysis.

We aimed to evaluate the segmentation performance of major organs within the imaging range of abdominal CT, one of the most common medical imaging modalities. To conduct this performance evaluation, we used a subset of the TotalSegmentator CT dataset version 1.0 [18]. The TotalSegmentator dataset is a large-scale, multiorgan segmentation dataset collected from eight institutions. We selected this dataset for its comprehensive organ segmentation masks and available institutional metadata for each case. These segmentation masks underwent expert verification to ensure high quality, making the dataset particularly suitable for our study.

Our study included cases that encompassed the abdominal region, while CT angiography scans were excluded from the analysis.

To ensure representation from all eight institutions while managing the dataset size, we implemented a sampling strategy. We set a maximum of 20 cases per institution and randomly selected cases up to this limit. For institutions with fewer than 20 cases, all available cases were included.

We focused on 8 major abdominal organs for our analysis:

These organs were selected based on their clinical significance and visibility in standard abdominal CT scans. To account for potential annotation deficiencies, we excluded segmentation masks with extremely small volumes by setting a threshold of 100 voxels. Masks below this threshold were omitted from the analysis. The dataset selection flowchart is illustrated in Figure 1.

The dataset was available in NIfTI file format. For SAM 2 inference, we extracted each horizontal slice from the 3D volumes to create subsets of 2D images for each scan. We applied windowing to the CT scans, using a window level of 50 and a window width of 400 Hounsfield units. Following windowing, we performed min-max scaling on the data. The scaled values were then converted to 8-bit integers, resulting in a range of 0‐255. These processed 2D images were saved as sequential JPEG files.

For SAM 2 inference, we selectively processed only the slices containing abdominal organs. This approach focused on optimizing computational efficiency, resulting in faster inference speeds.

For the volumetric analysis, we used the existing segmentation masks from the TotalSegmentator dataset to calculate the volume of each organ in voxels. We chose to measure in voxels rather than physical units, as our model inputs do not consider voxel scale.

Additionally, we analyzed cross-sectional areas of organ masks along the z-axis. For each organ, we calculated mean mask areas at the 25th, 50th, and 75th percentile z-coordinates, corresponding to the initial prompt locations used in SAM 2 segmentation.

SAM 2 is a segmentation model not specifically designed for medical images, but for general video content such as sports or animal footage. These models are trained on a large-scale dataset, enabling them to perform segmentation on any object. SAM 2’s main feature is its ability to support not only 2D images but also videos. By providing coordinates indicating the target object for segmentation, SAM 2 can track and segment objects appearing within the video. An overview of CT volume segmentation using SAM 2’s video predictor is shown in Figure 2.

To evaluate the model’s performance, we calculated the Dice similarity coefficient (DSC) [20]. This evaluation was performed organ-wise across the dataset to provide a detailed analysis.

We then compared the segmentation performance across the different approaches with or without negative prompts. We performed three pairwise comparisons for the approaches: caudal-approach versus mid-approach, caudal-approach versus cranial-approach, and mid-approach versus cranial-approach. Additionally, we compared performance with and without negative prompts.

When considering organ volumes in detail, we calculated Spearman correlation coefficients to examine the relationship between organ volumes and DSCs [21,22].

To account for multiple comparisons, we applied the Bonferroni correction. After Bonferroni correction, a P value of <.05/3 (approximately 0.0167) was considered statistically significant.

All statistical analyses were conducted using SciPy (version 1.14.0).

Our sampling strategy resulted in a total of 123 scans. Twenty scans each were selected from five institutions, while the remaining three institutions contributed 5, 5, and 13 cases respectively. The average age of the patients in our selected sample was 60.7 (SD 15.5) years. The gender distribution was nearly equal: 63 male and 60 female individuals.

903 organ segmentations were obtained from 123 scans. 12 masks with volumes of 100 voxels or smaller were then excluded from the analysis, and the final dataset consisted of 891 organ segmentations.

Organ volumes, measured in voxels and are detailed in Table 1. The liver was the largest organ, followed by the spleen; kidneys were next in size, with similar volumes on the left and right side. Compared to the liver’s mean volume, the pancreas was approximately 1/25th the size of the liver; the gallbladder was less than 1/70, and both adrenal glands were less than 1/400. These three organs (ie, pancreas, gallbladder, and adrenal glands) can be categorized as small organs.

Organ cross-sectional area analysis showed diverse trends across eight organs. The liver was the largest in size, increasing caudally to cranially. The pancreas steadily increased while adrenal glands, though smallest, peaked at mid-level. The details are provided in Figure S1 in Multimedia Appendix 1.

We evaluated the performance for multiorgan segmentation using different starting slice positions. The DSCs are reported as mean (SD) to reflect performance variability. All results are detailed in Table 2.

The left kidney demonstrated the best overall performance, maintaining high DSCs across all starting positions: mean 0.870 (SD 0.154) , mean 0.825 (0.221), and 0.808 (SD 0.242) for the caudal-approach, mid-approach, and cranial-approach, respectively. Notably, it was the only organ showing no statistically significant differences between any starting positions (all P>.0167).

The box plots (Figure 3) show that most organs reached DSCs above 0.8, with some approaching or nearly reaching 1.0. However, the box plots also reveal instances of very low DSC values approaching 0 across various organs and approaches, indicating significant variability in segmentation performance.

Starting slice level had a significant impact on most organs. Organs demonstrated various patterns in segmentation performance depending on the starting level. The liver showed a significant decrease in performance as the starting position moved superiorly, with DSC dropping from mean 0.821 (SD 0.192) with mean caudal-approach to 0.702 (SD 0.259) with cranial-approach (P<.01). Smaller organs such as the pancreas, adrenal glands, and gallbladder showed the most pronounced impact of the starting position. For these organs, performance significantly decreased when changing from a caudal-approach to a cranial-approach (all Ps<.01).

Larger organs such as liver, kidneys, and spleen consistently demonstrated higher DSCs compared to smaller organs across all approaches. A moderate correlation was observed across all settings when using caudal-approach (ρ=0.731; P<.01), mid-approach (ρ=0.698; P<.01), and cranial-approach (ρ=0.699; P<.01) (12,13). As shown in Table 3, when correlation coefficients were calculated separately for each organ, fair correlations were demonstrated for almost all items, particularly for smaller organs.

We also investigated the impact of including negative prompts on segmentation performance across different organs, focusing specifically on the caudal-approach (Table 4 and Figure S2 in Multimedia Appendix 1). All organs except the liver (P=.32) and spleen (P=.27) demonstrated significant increases in DSC (P<.01) with the inclusion of negative prompts.

In Figure 4, we present the highest DSC masks, excluding cases where the ground truth segmentations were incomplete. The highest performing masks, as visualized, generated for each organ in 3D were nearly indistinguishable from the ground truth.

On the other hand, there were cases where the performance fluctuated significantly due to differences in the approach. We present an example of the liver segmentation results in Figure 5. The DSC decreased by 0.564 (from 0.924 with caudal-approach to 0.360 with cranial-approach). For the caudal-approach, the initial slice segmentation appears to have been easier due to clear contrast with the surroundings. The cranial-approach resulted in lower DSC values compared to the caudal-approach. Visual inspection of the segmentation results showed incomplete masking and potential inclusion of the inferior vena cava in the liver mask, particularly in areas where boundaries between the liver, inferior vena cava, and abdominal wall were less distinct.