Hierarchical Map Research Articles

An efficient CNN accelerator for pattern-compressed sparse neural networks on FPGA

Currently, the sparsity of weights and activations are mainly utilized to improve the energy efficiency and computational performance of CNN accelerators. However, the irregular sparsity of weights and activations can lead to problems such as mapping access conflict and imbalanced workload, which poses great challenges to the efficient computation of sparse networks. To address those problems, this paper proposes a CNN accelerator for pattern-compressed sparse neural networks on FPGA to achieve efficient inference acceleration. It mainly includes three parts: first, we propose a convolutional kernel grouping fusion method to achieve a balanced workload of weights. Second, we propose an activation hash mapping conflict reduction method to drastically reduce the mapping access conflict rate. Finally, we further propose a hierarchical mapping computing architecture to achieve efficient processing of mapping access conflicts and balanced adaptation of weight loads. To verify the effectiveness of the method proposed in this paper, our accelerator is designed and implemented on the Zynq UltraScale+ MPSoC ZCU102 evaluation board and evaluated by running VGG16 and ResNet50 networks. The experimental results show that the method proposed in this paper can reduce the mapping conflicts by more than 97 % and improve the utilization rate of processing element (PE) by 1.67×. Furthermore, the accelerator implemented in this paper can achieve 393.2 GOPs throughput, 1.40–3.25× computational performance, and 1.11–7.34× energy efficiency improvement.

Exploring the customer perceived value of online grocery shopping: a cross-sectional study of Korean and Chinese consumers using Means-End Chain theory

Objectives: Despite the growing market share of online grocery shopping, there is a need to understand customer perceived value due to the ongoing advancements in information technology. This study explores the connections between attributes, consequences, and values. Additionally, it conducts a cross-country comparison of consumers’ online grocery shopping behaviors to gain a deeper understanding of consumer market segments and any potential variations among them. Methods: Data was collected through an online questionnaire survey conducted from May 1 to 15, 2024, targeting 400 consumers in Seoul, Korea, and Shanghai, China, who have experience with online grocery shopping. The survey utilized the Means-End Chain theory and association pattern technique hard laddering. Data collation and analysis were conducted using the IBM SPSS Statistics 28.0 program. The LadderUX software was employed to analyze the links between attributes, consequences, and values and create the consumer purchasing process’s implication matrix and hierarchical value map (HVM). Results: The study identified key attributes that influence online grocery shopping decisions, including delivery service, price, freshness, and quality. Korean consumers demonstrated a higher sensitivity to price (19.0%) and delivery service (17.0%). In contrast, Chinese consumers prioritized delivery service (15.0%) and after-sales service (14.8%). Commonly cited consequences included time saving (12.6% for Koreans, 11.3% for Chinese), whereas prevalent values encompassed convenience (36.8% for Koreans, 19.6% for Chinese) and economic value (26.6% for Koreans, 14.7% for Chinese). The HVM underscored these insights, highlighting diverse consumer preferences and country-specific nuances. Conclusions: The findings highlight the current state of online food consumption and consumers’ value systems, revealing variations among countries. These findings offer empirical insights that can be used to create customized global marketing strategies that resonate with various consumer preferences and market dynamics.

Kylin-V: An open-source package calculating the dynamic and spectroscopic properties of large systems.

Quantum dynamics simulation and computational spectroscopy serve as indispensable tools for the theoretical understanding of various fundamental physical and chemical processes, ranging from charge transfer to photochemical reactions. When simulating realistic systems, the primary challenge stems from the overwhelming number of degrees of freedom and the pronounced many-body correlations. Here, we present Kylin-V, an innovative quantum dynamics package designed for accurate and efficient simulations of dynamics and spectroscopic properties of vibronic Hamiltonians for molecular systems and their aggregates. Kylin-V supports various quantum dynamics and computational spectroscopy methods, such as time-dependent density matrix renormalization group and our recently proposed single-site and hierarchical mapping approaches, as well as vibrational heat-bath configuration interaction. In this paper, we introduce the methodologies implemented in Kylin-V and illustrate their performances through a diverse collection of numerical examples.

Crop health assessment through hierarchical fuzzy rule-based status maps

Precision agriculture is evolving toward a contemporary approach that involves multiple sensing techniques to monitor and enhance crop quality while minimizing losses and waste of no longer considered inexhaustible resources, such as soil and water supplies. To understand crop status, it is necessary to integrate data from heterogeneous sensors and employ advanced sensing devices that can assess crop and water status. This study presents a smart monitoring approach in agriculture, involving sensors that can be both stationary (such as soil moisture sensors) and mobile (such as sensor-equipped unmanned aerial vehicles). These sensors collect information from visual maps of crop production and water conditions, to comprehensively understand the crop area and spot any potential vegetation problems. A modular fuzzy control scheme has been designed to interpret spectral indices and vegetative parameters and, by applying fuzzy rules, return status maps about vegetation status. The rules are applied incrementally per a hierarchical design to correlate lower-level data (e.g., temperature, vegetation indices) with higher-level data (e.g., vapor pressure deficit) to robustly determine the vegetation status and the main parameters that have led to it. A case study was conducted, involving the collection of satellite images from artichoke crops in Salerno, Italy, to demonstrate the potential of incremental design and information integration in crop health monitoring. Subsequently, tests were conducted on vineyard regions of interest in Teano, Italy, to assess the efficacy of the framework in the assessment of plant status and water stress. Indeed, comparing the outcomes of our maps with those of cutting-edge machine learning (ML) semantic segmentation has indeed revealed a promising level of accuracy. Specifically, classification performance was compared to the output of conventional ML methods, demonstrating that our approach is consistent and achieves an accuracy of over 90% throughout various seasons of the year.

Land use hierarchical mapping for delineation of groundwater potential zones in the Ba river basin using satellite imagery and GIS technology

Abstract The evaluation of groundwater potential is crucial for the best use and replenishment of groundwater resources, as well as for the appropriate growth and administration of a region. To study groundwater potential, it is necessary to analyze the factors that directly affect this resource. In addition to the criteria that contribute to the formation of groundwater, such as rainfall, land cover, geology, slope, etc., land use types and their infiltration capacity are also the main factors affecting the groundwater prospect. In recent years, remote sensing and geographic information systems (GIS) have emerged as the most important technologies in the field of groundwater research, which aid in determining groundwater potential. This paper presents a method to generate the land use hierarchical map based on the level of impact on the formation of groundwater in the Ba river basin using GIS technology and remote sensing. According to the classification result, the three classes of groundwater potential zones including “low”, “moderate”, “high” occupy 1.51%, 77.43%, and 21.06% of the study area respectively. The findings can be validated with the actual groundwater level at some locations that are evenly distributed throughout the basin. The obtained results indicate that there is an agreement between groundwater potential levels at each actual location and those depicted on the land use hierarchical map. The methodology utilized in this study is reliable, accurate and can be used to generate hierarchical maps of additional parameters. This paper will be a useful document for groundwater potential zoning in the study area

"All You Need Is Love" a Social Network Approach to Understanding Attachment Networks in Adulthood.

This study examined five dimensions of attachment network structure in a large sample of adults (n = 930, 57% female) between 24 to 80 years of age. We employed a newly validated diagrammatic measure, the web-based hierarchical mapping technique (WHMT), to measure the attachment strength to participants' five closest relationships and the physical distance to and between network members. Our first aim was to replicate existing research on the composition of adult attachment networks, exploring variations in network patterns by age, romantic status, and parental status. Our second aim was to explore four new network dimensions, including physical distance to network members, hierarchical patterns, centrality, and density. The results replicated previous work on network composition, highlighting the pivotal role of romantic partners as primary attachment figures through adulthood. The analysis of the new network dimensions revealed a clear divide between adults in romantic relationships and those who are not. Compared to the single adults, the adults in romantic relationships were more hierarchical in their attachment preferences, reported lower emotional connection to friends and parents, and lived farther from their network, which was also more geographically dispersed. In other words, romantically involved adults put more of their attachment eggs in one basket. The results also showed that the older adults tended to live further away from their attachment network and had a more geographically dispersed network compared to the younger adults.

Multi-Tier Land Use and Land Cover Mapping Framework and Its Application in Urbanization Analysis in Three African Countries

The population of Africa is expected to rise to 2.5 billion by 2050, with more than 80% of this increase concentrated in cities. Africa’s anticipated population growth has serious implications for urban resource utilization and management, necessitating multi-level monitoring efforts that can inform planning and decision-making. Commonly, broad extent (e.g., country level) urban change analyses only examine a homogenous “developed” or “built-up” area, which may not capture patterns influenced by the heterogeneity of landscape features within urban areas. Contrarily, studies examining landscape heterogeneity at a finer resolution are typically limited in spatial extent (e.g., single city level). The goal of this study was to develop and test a hierarchical integrated mapping framework using globally available Earth Observation data (e.g., Landsat, Sentinel-2, Sentinel-1, and nightlight imagery) and accessible methodologies to produce national-level land use (LU) and urban-level land cover (LC) map products which may support a range of global and local monitoring and planning initiatives. We test our multi-tier methodology across three rapidly urbanizing African countries for the 2016–2020 period: Ethiopia, Nigeria, and South Africa. The initial output of our methodology includes annual national land use maps (Tier 1) for the purpose of delineating the dynamic boundaries of individual urban areas and monitoring national LU change. To complement Tier 1 LU maps, we detailed urban heterogeneity through LC classifications within urban areas (Tier 2) delineated using Tier 1 LU maps. Based on country-optimized sets of selected features that leverage spatial/texture and temporal dimensions of available data, we obtained an overall map accuracy of between 65 and 80% for Tier 1 maps and between 60 and 80% for Tier 2 maps, dependent on the evaluation country, although with consistent performance across study years providing a solid foundation for monitoring changes. We demonstrate the potential applications for our products through various analyses, including urbanization-driven LU change, and examine LC urban patterns across the three African study countries. While our findings allude to general differences in urban patterns across national scales, further analyses are needed to better understand the complex drivers behind urban LC configurations and their change patterns across different countries, city sizes, and rates of urbanization. Our multi-tier mapping framework is a viable strategy for producing harmonious, multi-level LULC products in developing countries using publicly available data and methodologies, which can serve as a basis for a wide range of informative and insightful monitoring analyses.

Vision-Based Situational Graphs Exploiting Fiducial Markers for the Integration of Semantic Entities

Situational Graphs (S-Graphs) merge geometric models of the environment generated by Simultaneous Localization and Mapping (SLAM) approaches with 3D scene graphs into a multi-layered jointly optimizable factor graph. As an advantage, S-Graphs not only offer a more comprehensive robotic situational awareness by combining geometric maps with diverse, hierarchically organized semantic entities and their topological relationships within one graph, but they also lead to improved performance of localization and mapping on the SLAM level by exploiting semantic information. In this paper, we introduce a vision-based version of S-Graphs where a conventional Visual SLAM (VSLAM) system is used for low-level feature tracking and mapping. In addition, the framework exploits the potential of fiducial markers (both visible and our recently introduced transparent or fully invisible markers) to encode comprehensive information about environments and the objects within them. The markers aid in identifying and mapping structural-level semantic entities, including walls and doors in the environment, with reliable poses in the global reference, subsequently establishing meaningful associations with higher-level entities, including corridors and rooms. However, in addition to including semantic entities, the semantic and geometric constraints imposed by the fiducial markers are also utilized to improve the reconstructed map’s quality and reduce localization errors. Experimental results on a real-world dataset collected using legged robots show that our framework excels in crafting a richer, multi-layered hierarchical map and enhances robot pose accuracy at the same time.

A Comparative Analysis of Machine Learning Algorithms for Identifying Cultural and Technological Groups in Archaeological Datasets through Clustering Analysis of Homogeneous Data

Machine learning algorithms have revolutionized data analysis by uncovering hidden patterns and structures. Clustering algorithms play a crucial role in organizing data into coherent groups. We focused on K-Means, hierarchical, and Self-Organizing Map (SOM) clustering algorithms for analyzing homogeneous datasets based on archaeological finds from the middle phase of Pre-Pottery B Neolithic in Southern Levant (10,500–9500 cal B.P.). We aimed to assess the repeatability of these algorithms in identifying patterns using quantitative and qualitative evaluation criteria. Thorough experimentation and statistical analysis revealed the pros and cons of each algorithm, enabling us to determine their appropriateness for various clustering scenarios and data types. Preliminary results showed that traditional K-Means may not capture datasets’ intricate relationships and uncertainties. The hierarchical technique provided a more probabilistic approach, and SOM excelled at maintaining high-dimensional data structures. Our research provides valuable insights into balancing repeatability and interpretability for algorithm selection and allows professionals to identify ideal clustering solutions.

Inheritance of Some Traits in Crosses between Hybrid Tea Roses and Old Garden Roses.

The limited knowledge about the inheritance of traits in roses makes the efficient development of rose varieties challenging. In order to achieve breeding goals, the inheritance of traits needs to be explored. Additionally, for the inheritance of a trait like scent, which remains a mystery, it is crucial to know the success of parental traits in transmitting them to the next generation. Understanding this allows for accurate parental selection, ensuring sustainability in meeting market demand and providing convenience to breeders. The aim of this study was to assess the success of cross-combinations between scented old garden roses and hybrid tea roses used in cut roses in transferring their existing traits, with the objective of achieving scented cut roses. The evaluated traits included recurrent blooming, flower stem length, flower diameter, petal number, scent, and bud length of both parents and progenies. The inheritance of these traits was evaluated through theoretical evaluations, including calculating heterosis and heterobeltiosis and determining narrow-sense heritability. The combinations and examined traits were assessed using a hierarchical clustering heat map. The results of this study indicated that flower stem length, flower diameter, petal number, and bud length traits had a moderate degree of narrow-sense heritability, suggesting the influence of non-additive genes on these traits. This study observed a low success rate in obtaining progenies with scent in cross combinations between cut roses and old garden roses, indicating the challenges in obtaining scented genotypes. The discrepancy between the observed phenotypic rates and the expected phenotypic and genotypic rates, according to Punnett squares, suggests that the examined traits could be controlled by polygenic genes. The progenies were observed to exhibit a greater resemblance to old garden roses than hybrid tea roses and did not meet the commercial quality standards for cut flowers. The significant negative heterosis observed in 65.12% (petal number) and 99.61% (flower diameter) of the progenies provides strong evidence of resemblance to old garden roses. Considering these findings, it is recommended to consider old garden roses as parents, taking into account their suitability for other breeding objectives.

A dual-track feature fusion model utilizing Group Shuffle Residual DeformNet and swin transformer for the classification of grape leaf diseases

Grape cultivation is important globally, contributing to the agricultural economy and providing diverse grape-based products. However, the susceptibility of grapes to disease poses a significant threat to yield and quality. Traditional disease identification methods demand expert knowledge, which limits scalability and efficiency. To address these limitations our research aims to design an automated deep learning approach for grape leaf disease detection. This research introduces a novel dual-track network for classifying grape leaf diseases, employing a combination of the Swin Transformer and Group Shuffle Residual DeformNet (GSRDN) tracks. The Swin Transformer track exploits shifted window techniques to construct hierarchical feature maps, enhancing global feature extraction. Simultaneously, the GSRDN track combines Group Shuffle Depthwise Residual block and Deformable Convolution block to extract local features with reduced computational complexity. The features from both tracks are concatenated and processed through Triplet Attention for cross-dimensional interaction. The proposed model achieved an accuracy of 98.6%, the precision, recall, and F1-score are recorded as 98.7%, 98.59%, and 98.64%, respectively as validated on a dataset containing grape leaf disease information from the PlantVillage dataset, demonstrating its potential for efficient grape disease classification.

Sustainable strategy for rehabilitating phosphate mining sites and valorisation of phosphate industry by-products and sludge using pistachio tree (Pistacia atlantica), false pepper (Schinus molle), and eucalyptus (Eucalyptus globulus) trees

The development of anthroposols has been proposed as a new environmentally friendly approach to ensuring the successful revegetation of phosphate mining sites. The phosphate industry's by-products, including phosphogypsum (PG), phosphate sludge (PS), and sewage sludge (SS), can be valuable resources in restoring the ecological balance of mined soil areas. The aim of this study was to safely and sustainably restore the ecological integrity of the phosphate mining site through the evaluation of nutrients and heavy metals dynamics in soil and plant tissues of three tree species and treated by-products containing 65 % PG, 30 % PS, and 5 % SS. The tree species used were Pistacia atlantica, Schinus molle, and Eucalyptus globulus. The experimental layout was a randomised complete block design with six replicates and three treatments. Growth diameter, height, nutrient uptakes and heavy metal dynamic were evaluated from the rhizosphere soils and plant tissues over two years. Hierarchical head maps of correlations between the measured growth parameters, soil and nutrient uptakes of the tree species were analysed using a phylogenetic generalised linear mixed model. S. molle and E. globulus had higher average diameter and height than P. atlantica plants. P. atlantica and S. molle showed greater nitrogen, phosphorus, potassium, calcium, and magnesium concentrations than E. globulus trees. Tree growth parameters were closely linked to soil nutrient bioavailability. The heavy metal accumulation ratio was higher in the E. globulus and S. molle leaves than in stems. Using by-products could be valorised for rehabilitating mine sites together with E. globulus and S. molle species.

International Skin Imaging Collaboration-Designated Diagnoses (ISIC-DX): Consensus terminology for lesion diagnostic labeling.

A common terminology for diagnosis is critically important for clinical communication, education, research and artificial intelligence. Prevailing lexicons are limited in fully representing skin neoplasms. To achieve expert consensus on diagnostic terms for skin neoplasms and their hierarchical mapping. Diagnostic terms were extracted from textbooks, publications and extant diagnostic codes. Terms were hierarchically mapped to super-categories (e.g. 'benign') and cellular/tissue-differentiation categories (e.g. 'melanocytic'), and appended with pertinent-modifiers and synonyms. These terms were evaluated using a modified-Delphi consensus approach. Experts from the International-Skin-Imaging-Collaboration (ISIC) were surveyed on agreement with terms and their hierarchical mapping; they could suggest modifying, deleting or adding terms. Consensus threshold was >75% for the initial rounds and >50% for the final round. Eighteen experts completed all Delphi rounds. Of 379 terms, 356 (94%) reached consensus in round one. Eleven of 226 (5%) benign-category terms, 6/140 (4%) malignant-category terms and 6/13 (46%) indeterminate-category terms did not reach initial agreement. Following three rounds, final consensus consisted of 362 terms mapped to 3 super-categories and 41 cellular/tissue-differentiation categories. We have created, agreed upon, and made public a taxonomy for skin neoplasms and their hierarchical mapping. Further study will be needed to evaluate the utility and completeness of the lexicon.

HierCGRA: A Novel Framework for Large-scale CGRA with Hierarchical Modeling and Automated Design Space Exploration

Coarse-grained reconfigurable arrays (CGRAs) are promising design choices in computation-intensive domains, since they can strike a balance between energy efficiency and flexibility. A typical CGRA comprises processing elements (PEs) that can execute operations in applications and interconnections between them. Nevertheless, most CGRAs suffer from the ineffectiveness of supporting flexible architecture design and solving large-scale mapping problems. To address these challenges, we introduce HierCGRA, a novel framework that integrates hierarchical CGRA modeling, Chisel-based Verilog generation, LLVM-based data flow graph (DFG) generation, DFG mapping, and design space exploration (DSE). With the graph homomorphism (GH) mapping algorithm, HierCGRA achieves a faster mapping speed and higher PE utilization rate compared with the existing state-of-the-art CGRA frameworks. The proposed hierarchical mapping strategy achieves 41× speedup on average compared with the ILP mapping algorithm in CGRA-ME. Furthermore, the automated DSE based on Bayesian optimization achieves a significant performance improvement by the heterogeneity of PEs and interconnections. With these features, HierCGRA enables the agile development for large-scale CGRA and accelerates the process of finding a better CGRA architecture.

Exploring the influence of attention for whole-image mammogram classification

Attention is an important component of modern Convolutional Neural Networks (CNNs) that has been shown to improve baseline model performance for a wide variety of tasks. Attention has shown specific promise in the classification and segmentation of mammograms, but we have a limited understanding of why attention improves performance in these domains. In this paper, we present a robust comparison of different combinations of baseline models and attention methods at two resolutions for whole mammogram classification of masses and calcifications. We find that attention generally helps to improve baseline model performance. However, the extent of improvement is governed by a combination of model architecture and the statistical characteristics of the data. Specifically, we show that high amounts of pooling and model complexity may result in decreased performance for data with high variability. To better understand the effect of attention on mammogram classification, we used LayerCAM, a hierarchical Class Activation Map (CAM) approach, to visualize where the network pays attention in the input image. This research provides statistical evidence that attention can improve the correlation between model performance and LayerCAM activation in the region of interest (ROI). However, these correlations are weak and variable, indicating that improvements in model performance due to attention are not necessarily caused by increased model activation near the ROI. Overall, our work provides novel insights to help guide future efforts in incorporating attention-based mechanisms for mammogram classification.

Hierarchical Mapping of Large-Scale Spiking Convolutional Neural Networks Onto Resource-Constrained Neuromorphic Processor

Hierarchical Mapping of Large-Scale Spiking Convolutional Neural Networks Onto Resource-Constrained Neuromorphic Processor

A small target detection algorithm based on improved YOLOv5 in aerial image.

Uncrewed aerial vehicle (UAV) aerial photography technology is widely used in both industrial and military sectors, but remote sensing for small target detection still faces several challenges. Firstly, the small size of targets increases the difficulty of detection and recognition. Secondly, complex aerial environmental conditions, such as lighting changes and background noise, significantly affect the quality of detection. Rapid and accurate identification of target categories is also a key issue, requiring improvements in detection speed and accuracy. This study proposes an improved remote sensing target detection algorithm based on the YOLOv5 architecture. In the YOLOv5s model, the Distribution Focal Loss function is introduced to accelerate the convergence speed of the network and enhance the network's focus on annotated data. Simultaneously, adjustments are made to the Cross Stage Partial (CSP) network structure, modifying the convolution kernel size, adding a new stack-separated convolution module, and designing a new attention mechanism to achieve effective feature fusion between different hierarchical structure feature maps. Experimental results demonstrate a significant performance improvement of the proposed algorithm on the RSOD dataset, with a 3.5% increase in detection accuracy compared to the original algorithm. These findings indicate that our algorithm effectively enhances the precision of remote sensing target detection and holds potential application prospects.

Theoretical investigation of distal charge separation in a perylenediimide trimer

An exciton–phonon (ex–ph) model based on our recently developed block interaction product basis framework is introduced to simulate the distal charge separation (CS) process in aggregated perylenediimide (PDI) trimer incorporating the quantum dynamic method, i.e., the time-dependent density matrix renormalization group. The electronic Hamiltonian in the ex–ph model is represented by nine constructed diabatic states, which include three local excited (LE) states and six charge transfer (CT) states from both the neighboring and distal chromophores. These diabatic states are automatically generated from the direct products of the leading localized neutral or ionic states of each chromophore’s reduced density matrix, which are obtained from ab initio quantum chemical calculation of the subsystem consisting of the targeted chromophore and its nearest neighbors, thus considering the interaction of the adjacent environment. In order to quantum-dynamically simulate the distal CS process with massive coupled vibrational modes in molecular aggregates, we used our recently proposed hierarchical mapping approach to renormalize these modes and truncate those vibrational modes that are not effectively coupled with electronic states accordingly. The simulation result demonstrates that the formation of the distal CS process undergoes an intermediate state of adjacent CT, i.e., starts from the LE states, passes through an adjacent CT state to generate the intermediates (∼200 fs), and then formalizes the targeted distal CS via further charge transference (∼1 ps). This finding agrees well with the results observed in the experiment, indicating that our scheme is capable of quantitatively investigating the CS process in a realistic aggregated PDI trimer and can also be potentially applied to exploring CS and other photoinduced processes in larger systems.

Interpretable multi-domain meta-transfer learning for few-shot fault diagnosis of rolling bearing under variable working conditions

To address the challenges of accurately diagnosing few-shot fault samples obtained from rolling bearings under variable operating conditions, as well as the issues of black box nature and delayed feedback to guide fault handling in intelligent diagnostic models, this paper proposes an interpretable multi-domain meta-transfer learning method. Firstly, vibration monitoring data of rolling bearings under different operating conditions are collected, and time–frequency domain features are extracted to construct multi-channel one-dimensional temporal samples as inputs. A multi-domain meta-transfer learning framework based on deep convolutional neural networks is then built to perform few-shot learning with multiple tasks under different operating conditions. The output results are reverse-reconstructed through a fusion hierarchical class activation mapping, and the feature maps are assigned different weights to obtain saliency maps corresponding to the inputs, thus improving the interpretability of the output results. Finally, the dataset of bearing vibration data under time-varying rotational speed conditions is used to validate the effectiveness of the proposed method. Experimental results show that the proposed method can achieve accurate fault diagnosis results under variable operating conditions with few-shot samples, and the diagnosis results can be fed back to the input for decision-making, enhancing the interpretability of the model. Compared with other models, it also demonstrates better robustness and accuracy.

DPA-P2PNet: Deformable Proposal-Aware P2PNet for Accurate Point-Based Cell Detection

Point-based cell detection (PCD), which pursues high-performance cell sensing under low-cost data annotation, has garnered increased attention in computational pathology community. Unlike mainstream PCD methods that rely on intermediate density map representations, the Point-to-Point network (P2PNet) has recently emerged as an end-to-end solution for PCD, demonstrating impressive cell detection accuracy and efficiency. Nevertheless, P2PNet is limited to decoding from a single-level feature map due to the scale-agnostic property of point proposals, which is insufficient to leverage multi-scale information. Moreover, the spatial distribution of pre-set point proposals is biased from that of cells, leading to inaccurate cell localization. To lift these limitations, we present DPA-P2PNet in this work. The proposed method directly extracts multi-scale features for decoding according to the coordinates of point proposals on hierarchical feature maps. On this basis, we further devise deformable point proposals to mitigate the positional bias between proposals and potential cells to promote cell localization. Inspired by practical pathological diagnosis that usually combines high-level tissue structure and low-level cell morphology for accurate cell classification, we propose a multi-field-of-view (mFoV) variant of DPA-P2PNet to accommodate additional large FoV images with tissue information as model input. Finally, we execute the first self-supervised pre-training on immunohistochemistry histopathology image data and evaluate the suitability of four representative self-supervised methods on the PCD task. Experimental results on three benchmarks and a large-scale and real-world interval dataset demonstrate the superiority of our proposed models over the state-of-the-art counterparts. Codes and pre-trained weights are available at https://github.com/windygoo/DPA-P2PNet.