Wei Dai

Accurate and robust cell detection in microscopic images is a fundamental yet challenging task due to diverse imaging conditions, dense cell distributions, and morphological variability. In this study, we propose FACellDet, a novel dual-branch hierarchical encoder-decoder framework that integrates adaptive feature alignment and a new supervisory signal to enhance cell detection. Specifically, a Coordinate-Attention-based Feature Alignment (CAFA) module is introduced to address spatial misalignment during multi-scale feature fusion, substantially improving cell detection precision. Furthermore, we design a Focal Attenuated Distance (FAD) map as an intermediate representation, providing highly discriminative and spatially informative cues, particularly in crowded regions. FACellDet features a dual-branch architecture, with the main branch predicting FAD maps for cell detection, while the auxiliary branch generates density maps to estimate cell counts for suppressing false detections. Extensive experiments on diverse cell types and imaging modalities from multiple public and in-house datasets demonstrate that our approach outperforms state-of-the-art methods in detection accuracy, while maintaining strong adaptability and robustness across challenging biomedical imaging scenarios. These results underscore the potential of FACellDet as an accurate and generalizable solution for automated cell detection in heterogeneous microscopic cell images, thereby facilitating reliable cell analysis to accelerate biomedical research and clinical workflows. Note to Practitioners—This work addresses the need for accurate and efficient cell detection and counting in biomedical images, where manual methods are time-consuming and error-prone, and existing automated approaches often struggle with dense or diverse cells. FACellDet offers a practical deep learning solution adaptable to various cell types and imaging conditions, improving both detection accuracy and robustness through innovative feature alignment and enhanced supervisory signals. This system can streamline laboratory workflows and support high-throughput research and clinical diagnostics. While FACellDet demonstrates strong performance across challenging datasets, its current deployment requires adequate computational resources. Future development could focus on creating lightweight versions and integrating the framework with automated imaging systems, further broadening its accessibility and impact in routine biomedical practice.

Early detection and accurate diagnosis can predict the risk of malignant disease transformation, thereby increasing the probability of effective treatment. Identifying mild syndrome with small pathological regions serves as an ominous warning and is fundamental in the early diagnosis of diseases. While deep learning algorithms, particularly convolutional neural networks (CNNs), have shown promise in segmenting medical objects, analyzing small areas in medical images remains challenging. This difficulty arises due to information losses and compression defects from convolutional and pooling operations in CNNs, which become more pronounced as the network deepens, especially for small medical objects. To address these challenges, we propose a novel scale-variant attention-based network (SvANet) for accurately segmenting small-scale objects in medical images. The SvANet consists of scale-variant attention (SvAttn), cross-scale guidance, Monte Carlo attention (MCAttn), and Vision Transformer (ViT), which incorporates cross-scale features and alleviates compression artifacts for enhancing the discrimination of small medical objects. Quantitative experimental results demonstrate the superior performance of SvANet, achieving 96.12%, 96.11%, 89.79%, 84.15%, 80.25%, 73.05%, and 72.58% in mean Dice (mDice) coefficient for segmenting kidney tumors, skin lesions, hepatic tumors, polyps, surgical excision cells, retinal vasculatures, and sperms, which occupy less than 1% of the image areas in KiTS23, ISIC 2018, ATLAS, PolypGen, TissueNet, FIVES, and SpermHealth datasets, respectively.

Designing millirobots capable of navigating high-friction environments remains a significant challenge due to limitations in force output and the absence of efficient transmission mechanisms at small scales. In this study, we introduce a magnetically inner actuated millirobot capable of generating a thrust force exceeding 15 N with a body weight of 5.82 g for moving across diverse frictional terrains. The inner-actuated millirobot features a dual-coil array positioned at each end of a plastic skeleton and a permanent magnet accommodated in the center channel of the body. When powered by a 0.5 A current, the internal magnetic interaction propels the magnet to a velocity of 2.10 m/s within 17 ms, striking the end wall to produce a powerful instantaneous thrust that overcomes friction forces. Experimental results demonstrate the millirobot’s ability to operate in viscous oil, traverse sand and granular media, and transport cargo exceeding 300 times its body weight. Furthermore, the magnetically inner actuated millirobot shows promising potential for accessing confined tubular environments. This magnetically inner-actuated design, leveraging momentum conservation for propulsion, enables millirobots with high force capacity for high-friction and confined-space applications.

Microscopic cell detection is a challenging task due to significant inter-cell occlusions in dense clusters and diverse cell morphologies. This paper introduces a novel framework designed to enhance automated cell detection. The proposed approach integrates a deep learning model that produces an inverse distance transform-based detection map from the given image, accompanied by a secondary network designed to regress a cell density map from the same input. The inverse distance transform-based map effectively highlights each cell instance in the densely populated areas, while the density map accurately estimates the total cell count in the image. Then, a custom counting-aided cell center extraction strategy leverages the cell count obtained by integrating over the density map to refine the detection process, significantly reducing false responses and thereby boosting overall accuracy. The proposed framework demonstrated superior performance with F-scores of 96.93%, 91.21%, and 92.00% on the VGG, MBM, and ADI datasets, respectively, surpassing existing state-of-the-art methods. It also achieved the lowest distance error, further validating the effectiveness of the proposed approach. These results demonstrate significant potential for automated cell analysis in biomedical applications.

An unbiased assessment of sperm morphology and motility is crucial for assessing fertility potential and guiding visual feedback for microrobotic manipulation. Automated analysis and selection of optimal sperm are essential for in vitro fertilization treatments, such as robotic intracytoplasmic sperm injection. However, conventional image processing methods face limitations in analyzing small sperm objects under microscopic imaging. While convolutional neural networks (CNNs) have brought promising advancements in microscopic image analysis, previous CNN methods have struggled to accurately differentiate tiny objects. These methods often require staining or fluorescence techniques to enhance visual contrast between sperm and culture medium, leading to clinical impracticality. To address these limitations, we introduce a novel sperm recognition network named the sperm feature-correlated network (SFCNet), for accurate and efficient segmentation and tracking of minute sperm objects. The SFCNet employs innovative modules, including collateral multi-scale convolution, cross-scale feature map guide, atrous spatial pyramid convolution with pooling, lateral attention, and multi-scale tracking proposal, to preserve essential sperm details despite their small size. Experimental results indicate that the SFCNet surpassed the state-of-the-art models designed for segmenting or tracking small objects, achieving up to a 28.39% higher Sorensen-Dice coefficient in segmentation and a 10.33% higher average precision in tracking. Additionally, the SFCNet excelled in sperm morphometric analysis, achieving errors below 15%. Moreover, the SFCNet also secured top-tier performance in sperm motility analysis, acquiring errors below 13% in seven sperm motility parameters. Note to Practitioners —This study is stimulated by the need to analyze the quality of motile sperms and select the optimal one for in vitro fertilization. Existing methods for detecting sperm fall short as they require a relatively high-magnification microscopic image or the usage of stain or fluorescence to increase sperm visualization, which limits the selection process or even makes the sperm clinically unavailable. To overcome these limitations, the present work proposes a new framework based on deep learning, which includes the design of extracting multi-scale sperm features. Experimental results suggest that the proposed method can perform better than existing methods in real-time analysis of multiple motile sperms’ morphology and motility at 20 $\times$ objective. In the future, there is a high potential for fertility specialists and healthcare workers to apply the presented framework in fertility treatment with higher accuracy and efficiency.

Cell counting is an essential step in a wide variety of biomedical applications, such as blood examination, semen assessment, and cancer diagnosis. However, microscopic cell counting is conventionally labor-intensive and error-prone for experts, and most of the existing automatic approaches are confined to a specific image type. To address these challenges, we propose a new interactive dual-network framework for automatic and generic cell counting. In this framework, one deep learning model (counter) is trained to regress a density map from a given microscope image. The number of cells in that image can be estimated by performing integration over the regressed density map. Another network (ground truth generator) is employed to dynamically generate suitable ground truth based on the cell samples and the dot annotations to serve as the supervision for training the counter. The interactive process to obtain the optimal model is achieved by jointly training the counter and ground truth generator iteratively. Moreover, we design a hierarchical multi-scale attention-based architecture to act as the counter in the proposed framework. This architecture is crafted to efficiently and effectively process multi-level features, enabling accurate regression of high-quality density maps. Evaluation experiments on three public cell counting datasets demonstrate the superiority of our method. Note to Practitioners —This paper is motivated by the need for advanced healthcare in the deep learning era. As a routine assessment procedure in healthcare settings, cell counting usually suffers from poor accuracy and inefficiency. We provide a solution to ameliorate the situation by developing a deep learning-based framework for automatic cell counting. After being trained in an end-to-end manner, the dual-network system is able to estimate the number of cells from the given microscopic images more accurately than existing methods. Additionally, this method is robust in various scenarios, such as calculating cell populations in suspension and cells in tissues. In the future, the presented pipeline has the potential to be implemented by biomedical practitioners who are non-expert in programming via wrapping it into a graphical user interface.

Accurate and unbiased examinations of skin lesions are critical for the early diagnosis and treatment of skin diseases. Visual features of skin lesions vary significantly because the images are collected from patients with different lesion colours and morphologies by using dissimilar imaging equipment. Recent studies have reported that ensembled convolutional neural networks (CNNs) are practical to classify the images for early diagnosis of skin disorders. However, the practical use of these ensembled CNNs is limited as these networks are heavyweight and inadequate for processing contextual information. Although lightweight networks (e.g., MobileNetV3 and EfficientNet) were developed to achieve parameter reduction for implementing deep neural networks on mobile devices, insufficient depth of feature representation restricts the performance. To address the existing limitations, we develop a new lite and effective neural network, namely HierAttn. The HierAttn applies a novel deep supervision strategy to learn the local and global features by using multi-stage and multi-branch attention mechanisms with only one training loss. The efficacy of HierAttn was evaluated by using the dermoscopy images dataset ISIC2019 and smartphone photos dataset PAD-UFES-20 (PAD2020). The experimental results show that HierAttn achieves the best accuracy and area under the curve (AUC) among the state-of-the-art lightweight networks. The code is available at https://github.com/anthonyweidai/HierAttn .