Page 129 - AIH-2-4

Artificial Intelligence in Health RefSAM3D for medical image segmentation

Ref-SAM3D achieved a dice score of 70.14%, marking a identified and segmented boundaries between different
10.11% increase over current technologies. It is noteworthy organs, maintaining high segmentation precision even
that traditional methods like nnU-Net perform well on in cases with blurred organ boundaries or complex
certain tasks, yet overall, they fall short compared to newer anatomical structures. Notably, Ref-SAM3D maintained
methods such as Ref-SAM3D. Particularly, when dealing stable segmentation performance for both small organs
with tumors that have blurred boundaries and diverse such as the pancreas, and elongated structures, such as the
morphologies, Ref-SAM3D demonstrated significant aorta, further validating the reliability of the quantitative
advantages. These findings underscore the exceptional evaluation metrics.
performance of Ref-SAM3D in addressing a variety of In addition, in the context of cardiac tumor segmentation
complex medical image segmentation challenges. Figure 3 using MRI, as shown in Figure 5, a qualitative assessment
shows the qualitative visualizations of these tasks. of predicted masks from various segmentation models
In the domain of multi-organ segmentation, we indicates that our AutoSAM Adapter produced visually
conducted experiments on the BTCV dataset. The Ref- superior results, especially in terms of boundary precision,
SAM3D approach demonstrated exceptional capability, when compared to existing state-of-the-art methods.
achieving a dice score of 97.1% for spleen segmentation, 4.3. Generalization evaluation
as shown in Table 3, which surpasses all comparative
methods. The left and right kidneys attained dice scores To assess the generalization capabilities of Ref-SAM3D,
of 96.1% and 94.9%, respectively. The esophagus achieved we conducted comprehensive experiments across
a dice score of 85.2%, surpassing other methods, whereas heterogeneous datasets and imaging modalities. Our
the liver and stomach achieved scores of 97.3% and 94.1%, evaluation framework encompassed two distinct scenarios:
respectively. Furthermore, Ref-SAM3D demonstrated cross-modality generalization on the AMOS22 dataset
efficiency in handling complex anatomical structures, such (comprising both CT and MRI modalities) and cross-
as the pancreas and aorta, achieving dice scores of 87.5% dataset adaptation using the MM-WHS cardiac imaging
and 92.3%, respectively. Ref-SAM3D achieved an average dataset.
HD value of 2.34, underscoring its superior boundary In the zero-shot generalization experiments, we
precision. Figure 4 shows qualitative visualizations of evaluated the model’s transferability by applying our Ref-
BTCV tasks. From the qualitative visualization results, SAM3D, trained exclusively on the BTCV CT dataset, to
Ref-SAM3D demonstrated superior performance in the AMOS22 dataset without any additional fine-tuning.
multi-organ segmentation tasks. The method accurately The quantitative results demonstrated remarkable

Figure 3. Qualitative visualizations of the proposed method and baseline approaches on liver tumor, kidney tumor, pancreas tumor, and colon cancer
segmentation tasks
Abbreviations: 3D: Three-dimensional; nn: No new; NSD: Normalized surface Dice; SAM: Segment Anything Model; UNETR: U-Net Transformers;
UX-Net: UNet-eXpanded Network

Volume 2 Issue 4 (2025) 123 doi: 10.36922/AIH025080010