Three-dimensional (3D) medical image segmentation typically demands extensive labeled training samples, which is prohibitively time-consuming and requires significant expertise. Although this demand can be mitigated by special learning paradigms such as semi-supervised learning, the cost is still high due to the reader-unfriendly 3D data structure. In this paper, we seek a solution of robust 3D segmentation using extremely simplified annotation that delineates only a single slice per each volume for only a subset of the 3D samples. To this end, we propose two innovative modules: a correlation-enhanced 3D segmentation model (CE-Seg) and a hybrid 3D pseudo mask generator (Hy-Gen). CE-Seg aims to comprehensively understand the 3D targets under super-sparse single-slice supervision by maximizing its ability to mine correlations across slices, spaces and scales. Specifically, CE-Seg mimics the radiologist's interpretation by 'seeing' a dynamically scrolling 3D image to enrich the slice-correlated context. It also introduces a drop-then-restoration self-played task to enhance the spatial correlations of features, and uses a bidirectional cascaded attention to interactively fuse features across different scales. To train CS-Seg, Hy-Gen combines learning-based and learning-free strategies to generate reliable pseudo 3D masks as supervisions. Concretely, Hy-Gen first employs a level-set evolution to 'spread' the single annotation to its neighboring slices as initialization. It then builds a teacher-student framework to progressively refine the initialized 3D mask by dynamically merging the predictions of the CS-Seg's teacher-copy. Extensive experiments on three public and one in-house datasets indicate that our method exceeds eight state-of-the-art semi-supervised methods by at least 3$\%$ in dice, and is even on par with the full-supervised counterpart.

The vessel-wall-volume (VWV) measured based on three-dimensional (3D) carotid artery (CA) ultrasound (US) images can help to assess carotid atherosclerosis and manage patients at risk for stroke. Manual involvement for measurement work is subjective and requires well-trained operators, and fully automatic measurement tools are not yet available. Thereby, we proposed a fully automatic VWV measurement framework (Auto-VWV) using a CA prior-knowledge embedded U-Net (CAP-UNet) to measure the VWV from 3D CA US images without manual intervention. The Auto-VWV framework is designed to improve the repeated VWV measuring consistency, which resulted in the first fully automatic framework for VWV measurement. CAP-UNet is developed to improve segmentation accuracy on the whole CA, which composed of a U-Net type backbone and three additional prior-knowledge learning modules. Specifically, a continuity learning module is used to learn the spatial continuity of the arteries in a sequence of image slices. A voxel evolution learning module was designed to learn the evolution of the artery in adjacent slices, and a topology learning module was used to learn the unique topology of the carotid artery. In two 3D CA US datasets, CAP-UNet architecture achieved state-of-the-art performance compared to eight competing models. Furthermore, CAP-UNet-based Auto-VWV achieved better accuracy and consistency than Auto-VWV based on competing models in the simulated repeated measurement. Finally, using 10 pairs of real repeatedly scanned samples, Auto-VWV achieved better VWV measurement reproducibility than intra- and inter-operator manual measurements. The code is available at https://github.com/Yue9603/Auto-VWV.

Semi-supervised segmentation is highly significant in 3D medical image segmentation. The typical solutions adopt a teacher-student dual-model architecture, and they constrain the two models' decision consistency on the same segmentation task. However, the scarcity of medical samples can lower the diversity of tasks, reducing the effectiveness of consistency constraint. The issue can further worsen as the weights of the models gradually become synchronized. In this work, we have proposed to construct diverse joint-tasks using masked image modelling for enhancing the reliability of the consistency constraint, and develop a novel architecture consisting of a single teacher but multiple students to enjoy the additional knowledge decoupled from the synchronized weights. Specifically, the teacher and student models 'see' varied randomly-masked versions of an input, and are trained to segment the same targets but reconstruct different missing regions concurrently. Such joint-task of segmentation and reconstruction can have the two learners capture related but complementary features to derive instructive knowledge when constraining their consistency. Moreover, two extra students join the original one to perform an inter-student learning. The three students share the same encoding but different decoding designs, and learn decoupled knowledge by constraining their mutual consistencies, preventing themselves from suboptimally converging to the biased predictions of the dictatorial teacher. Experimental on four medical datasets show that our approach performs better than six mainstream semi-supervised methods. Particularly, our approach achieves at least 0.61% and 0.36% higher Dice and Jaccard values, respectively, than the most competitive approach on our in-house dataset. The code will be released at https://github.com/zxmboshi/DDL.

In the field of disease diagnosis where only a small dataset of medical images may be accessible, the light-weight convolutional neural network (CNN) has become popular because it can help to avoid the over-fitting problem and improve computational efficiency. However, the feature extraction capability of the light-weight CNN is inferior to that of the heavy-weight counterpart. Although the attention mechanism provides a feasible solution to this problem, the existing attention modules, such as the squeeze and excitation module and the convolutional block attention module, have insufficient non-linearity, thereby influencing the ability of the light-weight CNN to discover the key features. To address this issue, we have proposed a spiking cortical model based global and local (SCM-GL) attention module. The SCM-GL module analyzes the input feature maps in parallel and decomposes each map into several components according to the relation between pixels and their neighbors. The components are weighted summed to obtain a local mask. Besides, a global mask is produced by discovering the correlation between the distant pixels in the feature map. The final attention mask is generated by combining the local and global masks, and it is multiplied by the original map so that the important components can be highlighted to facilitate accurate disease diagnosis. To appreciate the performance of the SCM-GL module, this module and some mainstream attention modules have been embedded into the popular light-weight CNN models for comparison. Experiments on the classification of brain MR, chest X-ray, and osteosarcoma image datasets demonstrate that the SCM-GL module can significantly improve the classification performance of the evaluated light-weight CNN models by enhancing the ability of discovering the suspected lesions and it is generally superior to state-of-the-art attention modules in terms of accuracy, recall, specificity and F1 score.