A novel MOSnet model for multi-object segmentation of medical images in micro-computed tomography

Datasets

Due to the widespread use of mouse model in Micro-CT research, in this study, two datasets were utilized to evaluate the segmentation accuracy of the proposed MOSnet: one native mouse Micro-CT dataset without any contrast agents and one contrast-enhanced mouse Micro-CT dataset with a contrast agent injected. All of them were obtained from Nature Scientific Data [27].

The native dataset consists of 140 whole-body scans from 20 mice by a preclinical Micro-CT. In this dataset, the field of view was 40.32 × 28.84 × 55.44 mm, and the voxel size was 0.28 × 0.28 × 0.28 mm [27]. The native dataset has the ground truth of these organs: heart, lungs, liver, kidneys, spleen, bladder and intestine.

The contrast-enhanced dataset involves 85 whole-body scans from 10 mice by a preclinical Micro-CT. Before the first scan, the contrast agent was intravenously injected, which is able to enhance the contrast of spleen and liver in Micro-CT. In the contrast-enhanced dataset, the field of view was 43.12 × 33.88 × 67.76 mm, and the voxel size was 0.28 × 0.28 × 0.28 mm [27]. This dataset has the ground truth of these organs: heart, lungs, liver, kidneys, spleen, bladder, intestine, stomach and bone.

Due to the different field of view and different size between the two datasets, 2D segmentation is more suitable than 3D segmentation [28]. Before the training, we preprocessed the data by adjusting the window width to 1000 HU and the window level to 400 HU for each slice, followed by normalization to a range of 0–1 to mitigate relative numerical discrepancies across the data. In order to achieve multi-organ segmentation, we trained both datasets in the same network simultaneously. A total of 50,591 slices were randomly divided at a ratio of 9:1 for training and validation respectively.

Evaluation metrics

In this study, we evaluated the performance of the proposed MOSnet quantitatively through three widely recognized evaluation metrics: dice similarity coefficient (DSC), jaccard similarity coefficient (JSC), and 95%hausdorff distance (HD95). Their calculation formulas are as follows:

where Y ˆ is the predicted result and Y is the binary mask. The higher scores of DSC and JSC indicate a higher segmentation accuracy because they both assess overlapping area between the predicted result and the ground truth. The HD95 calculates the maximum distance from each point in the predicted result to the nearest point in the binary mask, so its lower score indicates superior boundary segmentation results.

Implementation

In this study, we chose the Adam optimizer algorithm, setting 0.001 as the initial learning rate for updating the parameters of the network and a scheduler for updating the learning rate. The training is through 100 epochs using a batch-size of 16 images. All experiments were executed using PyTorch on a machine equipped with a GPU of NVIDIA 2080Ti.

Experiment results

For illustrating the efficacy of the proposed MOSnet, we compared its prediction results of the two datasets with those of four excellent segmentation networks: ResUnet [29], Unet3+ [30], DAVnet3+ [31], and AIMOS [25]. To maintain fairness in comparison, we chose identical preprocessing for these networks and refrained from postprocessing. Also, their batch-sizes, epochs, optimizers and schedulers are consistent for all networks.

The results of qualitative comparison are shown in Figure 2. As evident from these figures, despite these networks achieving high accuracy, the results of MOSnet are more consistent with the ground truth than the other networks. For the native dataset, our method accurately delineates the target regions of the liver, spleen, bladder, and intestine, exhibiting precise boundaries. Similarly, for the contrast-enhanced dataset, the stomach region is accurately portrayed with meticulous contours by our model. Although MOSnet outperforms other networks, there may be some areas where the segmentation results are not entirely satisfactory, particularly when dealing with lungs of the native dataset. Nevertheless, it consistently demonstrates a lower number of segmentation errors compared to other networks.

Figure 2:

Segmentation results of the five networks for the native dataset (UP) and the contrast-enhanced dataset (Down).

Tables 1 and 2 present the quantitative results of all methods evaluated on the two datasets. By analyzing the experimental data, the results of the proposed MOSnet exhibit its superior performance to ResUnet, Unet3+, DAVnet3+, and AIMOS. It can be observed that MOSnet, compared to the best-performing network, evenly improves DSC by 4.1 and 2.4 %, increased JSC by 4.1 and 3.1 %, simultaneously reducing HD95 by 16.3 and 19.3 % on the two datasets respectively. From the perspective of the segmentation results of each organ, the difficulty of segmenting intestine is the biggest, but our model still achieved best, particularly improving the DSC value by 0.078, 0.053, 0.014, 0.023 and 0.039, 0.064, 0.022, 0.009. This indicates that in the multi-organ segmentation task, MOSnet effectively enhances the accuracy especially in complex regions when completing the segmentation of various targets.

Additionally, Figure 3 displays several histograms about the three metrics of these five networks. It is obvious that no matter for the native dataset or the contrast-enhanced dataset, our proposed MOSnet has the best results on the multi-organ segmentation with the highest average DSC and JSC, and the lowest average HD95. Although the standard error of MOSnet on the contrast-enhanced dataset is somewhat higher partially compared to other segmentation models, it is exhibited the lowest across almost all organs on the native dataset. This observation suggests that our model possesses greater robustness and stability, especially when dealing with different datasets and multi-organ segmentation.

Figure 3:

Histograms of DSC, JSC, and HD95 of the networks (The first column represents the native dataset, and the second column represents the contrast-enhanced dataset. The colorful bars are the average values for different organs, and the black lines are the standard errors for the corresponding average values.)

Advantages of the proposed MOSnet

According to the above results of experiments for segmenting multiple organs, it has been suggested that our proposed MOSnet can accomplish high accuracy with training two datasets simultaneously, which jointly learned features of organs under different parameters. Furthermore, despite the incomplete labeling of the first dataset, the participation of the second dataset through collaborative learning enabled the segmentation of organs unlabeled in the first dataset.

To validate the ability of transfer learning employed by MOSnet, we conducted the following experiment: the labels of the kidneys and bladder in the second dataset were excluded from training and only used to calculate evaluation indicators. The results of this experiment are presented in Table 3 and Figure 4(a). By this way, we aimed to assess the model’s ability to generalize and adapt to untrained data. This experiment served as a rigorous test of MOSnet’s transfer learning capabilities, as it challenged the model to learn and segment organs effectively even without direct supervision from labeled data.

Figure 4:

The results of different experiments. (a) The results of kidneys and bladder. (b) Visual results of ablation study. (c) The results under different λ values. (d) The result of the reconstruction decoder.

The results presented in Table 3 and Figure 4(a) demonstrate the model can achieve precise segmentation of target organs by leveraging features learned from other datasets – even when partial organ labels are missing in a single dataset. This further validates the model’s ​​generalization capability​​ and the effectiveness of its ​​transfer learning​​ mechanism.

The results presented in Table 3 demonstrate the model can segment organs accurately enough even in the absence of labeled data, which demonstrated the power of MOSnet in handling various datasets and addressing challenges associated with incomplete annotations. This validates the effectiveness of MOSnet’s transfer learning approach, which leverages knowledge learned from one dataset to enhance the segmentation of organs in another dataset. The experiment highlights the significant potential of MOSnet in real-world scenarios where complete and accurate annotations may not always be available.

This capability was achieved by incorporating a control module and a reconstruction block, and we conducted several systematic experiments to evaluate the efficacy of these modules. The basic model (BM) of MOSnet adopts an end-to-end framework, which includes nested and dense skip connections. Based on BM, a control module and a reconstruction block were added to form the proposed method MOSnet. So, we compared MOSnet to the BM, the basic model with a control module (BM_CM) and the basic model with a reconstruction block (BM_RB) on the validation of the contrast-enhanced dataset.

The ablation study results and statistical significance testing results, shown in Table 4 and Figure 4(b), demonstrate that the BM_CM significantly enhances the performance of multi-organ segmentation compared to the BM. For instance, the stomach’s DSC improves from 0.873 (BM) to 0.903 (BM_CM), and the spleen’s JSC increases from 0.824 (BM) to 0.841 (BM_CM). These improvements indicate that the control module effectively refines the segmentation process by dynamically studying feature representations from different datasets. However, the BM_CM remains susceptible to incomplete segmentation, particularly for organs with ambiguous boundaries, such as the liver and intestine. For example, the liver’s DSC in BM_CM is 0.894, but our model achieves 0.937, a +4.8 % improvement, suggesting that the control module may over-suppress weak signals at the liver’s periphery, leading to information loss. Similarly, the intestine’s DSC in BM_CM is 0.766, significantly lower than the 0.836 in Ours. To address these limitations, a reconstruction block was integrated into the framework, forming a multi-task learning approach that generates reconstruction maps alongside segmentation probability maps. This design enables the model to retain fine anatomical details. It can be clearly seen from the performance of the complete model (our model): Among all organs, our model achieved the highest DSC/JSC value and the lowest HD95 value, and made almost all statistically significant improvements (p<0.05). For instance, the lungs’ HD95 decreases from 3.170 (BM_CM) to 1.897 (Ours), indicating sharper boundary localization, while the spleen’s JSC increases from 0.841 (BM_CM) to 0.927 (Ours), resolving ambiguities with adjacent organs. The reconstruction block enhances performance by preserving high-frequency details and implicitly regularizing the model through input reconstruction, reducing overfitting to noisy annotations.

In Equation (9), the weighting factor λ for the reconstruction task influences both the feature extraction capability of the shared encoder during training and the accuracy of the primary segmentation task. To optimize segmentation performance, we examined the variation in average DSC across all segmented organs under different λ values. As shown in Figure 4(c), when λ=0.7, the model achieves consistently higher DSC values for both datasets compared to other λ values. When λ<0.6, the model becomes overly focused on segmentation, diminishing the self-supervised regularization effect and resulting in marginal degradation in boundary precision. When λ>0.8, the reconstruction loss is overly down-weighted, weakening its role in preventing feature overfitting. This leads to slightly noisier predictions and reduced robustness. At λ=0.7, we observe the most consistent performance across organs and anatomical scales, suggesting it provides an effective trade-off for the tasks and data distribution in our study. Consequently, the proposed model adopts λ=0.7 in its overall loss function.

Moreover, as shown in Figure 4(d), these visualizations demonstrate that our model is able to generate anatomically plausible and structurally consistent reconstructions, indicating that the shared encoder has learned meaningful and semantically rich representations from the input micro-CT, supporting its utility as an auxiliary task.

Limitations and future works

The segmentation of mouse micro-CT scans plays a crucial role in biomedical research, as the accuracy of organ delineation directly affects the quality of s quantitative analysis. Compared to existing methods, our proposed MOSnet achieves superior segmentation performance across two datasets, primarily attributed to the effective design of the control module and reconstruction block. Nevertheless, several limitations warrant further discussion.

Firstly, although MOSnet demonstrates strong overall performance, segmentation accuracy for anatomically complex organs, particularly the spleen and intestines, still requires improvement. These structures often exhibit high variability in shape and intensity, posing challenges for precise boundary delineation. To enhance accuracy, future work will explore 3D adaptation and interpolation strategies that can fully exploit spatial contextual information across adjacent slices, thereby improving inter-slice coherence and structural consistency. Also, recent advances in vision transformers (ViTs) and hybrid transformer-convolutional architectures have demonstrated remarkable performance in medical image segmentation, particularly in capturing long-range dependencies and modeling global context – capabilities that are highly valuable for whole-body, multi-organ segmentation tasks. In ongoing and future work, we are actively exploring the integration of vision transformer architectures into our 3D segmentation framework for mouse micro-CT data. Secondly, as shown in Table 3, the model’s performance exhibits a noticeable decline when evaluated on external datasets, indicating limited generalization under domain shifts. This underscores the need to incorporate more diverse, multi-source datasets in future training. Moreover, preprocessing strategies such as style standardization or domain adaptation techniques will be essential to harmonize variations in imaging protocols, scanner characteristics, and noise profiles, thereby enhancing the model’s robustness and cross-domain applicability. It is important to note that all reported results in this study were obtained without any post-processing to ensure a fair comparison of the intrinsic segmentation performance of the different models. However, in practical clinical deployment, simple and computationally efficient post-processing techniques are commonly employed to refine segmentation. Lastly, while the current architecture is effective, its reliance on the control module and reconstruction block results in relatively high complexity, leading to increased computational cost and longer training times. Our model has 11.4M parameters, comparable to UNet3+ (9.2M). Training on a GPU of NVIDIA 2080Ti, MOSnet takes approximately 24 h on the full dataset, similar to UNet3+ (∼21.5 h) and 1.3× longer than ResUnet (∼18 h). To facilitate broader adoption and potential deployment in resource-constrained settings, we will focus on developing a more lightweight variant of MOSnet, aiming to maintain high performance while significantly improving efficiency. For instance, morphological operations such as closing or small object removal could potentially further improve the final segmentation quality.

In summary, future efforts will concentrate on integrating 3D contextual modeling, enhancing generalization through domain-invariant preprocessing, and designing a more lightweight network architecture – steps critical for advancing MOSnet from a research prototype toward practical application in preclinical imaging.