Yu Liang

The Segment Anything Model (SAM) demonstrates exceptional zero‐shot capabilities on natural images, yet its application to three‐dimensional medical images faces two major challenges: sensitivity to blurred boundaries and the difficulty of two‐dimensional architectures in effectively modeling three‐dimensional volumetric contextual information. To address this, we propose a medical segmentation method based on SAM (LSSM‐SAM), which completes automatic segmentation without any interactive prompts. This approach employs a parameter‐efficient fine‐tuning strategy, introducing a residual compensation mechanism to overcome the inherent limitations of the FacT method. It preserves the majority of SAM's pre‐trained weights while incrementally updating only a small portion. By introducing custom SC‐adapters within the image encoder, SAM's 2D backbone network adapts to extract features that integrate the third dimension. To prevent local feature loss during fine‐tuning, the FNN in the image decoder's Transformer is replaced with the novel state space model LSSM. The trained LSSM‐SAM model demonstrates superior performance across three datasets, surpassing multiple current state‐of‐the‐art methods.

In the field of micromanipulation, accurate identification and localization of micro-targets are crucial. This article presents a robust high-precision micromanipulation targets localization and motion tracking algorithm to address the challenges posed by uneven illumination, shadows, object occlusion, and image noise in micromanipulation. The proposed method employs improved bilateral filtering and the Sobel operator for illumination correction and edge contour detection in the horizontal view, achieving identification and Z-directional localization of micro-targets. In the vertical view, the improved grayscale template and the feature point matching algorithms accurately identify micro-targets and locate manipulation points in the X/Y directions, even under occlusion and rotation. Experimental results demonstrate that the algorithm can identify and localize micro-targets with an average error less than 1 μm and an average time of 120 ms, significantly improving the success rate, accuracy, and efficiency compared to existing methods. This article provides a novel approach for the precise automation of micromanipulation tasks.

Images at the micrometer level usually have high resolution and contain a large amount of detailed information, and traditional vision algorithms are designed for macroscopic images, making it difficult to achieve accurate target localization at the microscopic scale. In this paper, we propose a micro-target localization algorithm based on improved local contour extraction and feature point matching to address the problems of low accuracy and time-consuming operation point localization under microscopic vision due to uneven illumination, angular shift of micro-targets, and occlusion. In the horizontal perspective, a light source correction algorithm based on the morphological algorithm and an edge enhancement algorithm based on Fourier transform is proposed to improve the accuracy of threshold segmentation and edge extraction, and a contour feature extraction algorithm based on Normalized Cross-Correlation template matching and improved Otsu’s thresholding method is utilized to achieve high-precision localization of multi-targets in micro-scale. In the vertical perspective, a Binary Robust Invariant Scalable Keypoints matching algorithm based on spatial feature screening is proposed to solve the problems of feature point mismatch and inaccurate localization of traditional algorithms in case of angular offset and occlusion of micro-targets. Finally, experiments were conducted on the microscopic vision operating system and experimentally compared with cutting-edge methods to verify the feasibility and superiority of the present method. The experimental results show that the proposed algorithm in this paper has an average error of 1.023 pixels and an average elapsed time of 109.08 ms, exhibits higher stability in the presence of light source interference, angular offset, and occlusion of micro-targets, and significantly improves both localization accuracy and efficiency.

Gliomas are the most common type of primary brain tumors, classified into low‐grade gliomas (LGGs) and high‐grade gliomas (HGGs). There is a significant difference in survival rates between patients with different grades of gliomas, making imaging‐based grading a research hotspot. Current deep learning–based glioma grading algorithms face challenges, such as network complexity, low accuracy, and difficulty in large‐scale application. This paper proposes a multimodal, lightweight Xception grading network to address these issues. The network introduces convolutional block attention modules and employs dilated convolutions for spatial feature aggregation, reducing parameter count while maintaining the same receptive field. By integrating spatial and channel squeeze‐and‐excitation modules, it achieves more accurate feature learning, alongside improvements to the residual connection modules for critical information retention. Compared to existing methods, the proposed approach improves classification accuracy while maintaining a reduced parameter count. The network was trained and validated on 344 glioma cases (261 HGGs and 83 LGGs) and tested on 38 glioma cases (29 HGGs and 9 LGGs). Experimental results demonstrate that the network achieves an accuracy of 92.67% and an AUC of 0.9413 using a fully connected layer as the classifier. The features extracted using the improved Xception grading network achieved an accuracy of 93.42% when classified with KNN and RF classifiers. This study aims to provide diagnostic suggestions for clinical use through a simple, effective, and noninvasive multimodal medical imaging diagnostic method for LGG/HGG grading, thereby accelerating treatment decision‐making.