Recent work has shown that convolutional networks can be substantially deeper, more accurate, and efficient to train if they contain shorter connections between layers close to the input and those close to the output. In this paper, we embrace this observation and introduce the Dense Convolutional Network (DenseNet), which connects each layer to every other layer in a feed-forward fashion. Whereas traditional convolutional networks with L layers have L connections—one between each layer and its subsequent layer—our network has L(L+1)/2 direct connections. For each layer, the feature-maps of all preceding layers are used as inputs, and its own feature-maps are used as inputs into all subsequent layers. DenseNets have several compelling advantages: they alleviate the vanishing-gradient problem, strengthen feature propagation, encourage feature reuse, and substantially reduce the number of parameters. We evaluate our proposed architecture on four highly competitive object recognition benchmark tasks (CIFAR-10, CIFAR-100, SVHN, and ImageNet). DenseNets obtain significant improvements over the state-of-the-art on most of them, whilst requiring less memory and computation to achieve high performance. Code and pre-trained models are available at https://github.com/liuzhuang13/DenseNet.
We present a method for detecting objects in images using a single deep neural network. Our approach, named SSD, discretizes the output space of bounding boxes into a set of default boxes over different aspect ratios and scales per feature map location. At prediction time, the network generates scores for the presence of each object category in each default box and produces adjustments to the box to better match the object shape. Additionally, the network combines predictions from multiple feature maps with different resolutions to naturally handle objects of various sizes. Our SSD model is simple relative to methods that require object proposals because it completely eliminates proposal generation and subsequent pixel or feature resampling stage and encapsulates all computation in a single network. This makes SSD easy to train and straightforward to integrate into systems that require a detection component. Experimental results on the PASCAL VOC, MS COCO, and ILSVRC datasets confirm that SSD has comparable accuracy to methods that utilize an additional object proposal step and is much faster, while providing a unified framework for both training and inference. Compared to other single stage methods, SSD has much better accuracy, even with a smaller input image size. For $300\times 300$ input, SSD achieves 72.1% mAP on VOC2007 test at 58 FPS on a Nvidia Titan X and for $500\times 500$ input, SSD achieves 75.1% mAP, outperforming a comparable state of the art Faster R-CNN model. Code is available at this https URL .
We present a novel and practical deep fully convolutional neural network architecture for semantic pixel-wise segmentation termed SegNet. This core trainable segmentation engine consists of an encoder network, a corresponding decoder network followed by a pixel-wise classification layer. The architecture of the encoder network is topologically identical to the 13 convolutional layers in the VGG16 network [1] . The role of the decoder network is to map the low resolution encoder feature maps to full input resolution feature maps for pixel-wise classification. The novelty of SegNet lies is in the manner in which the decoder upsamples its lower resolution input feature map(s). Specifically, the decoder uses pooling indices computed in the max-pooling step of the corresponding encoder to perform non-linear upsampling. This eliminates the need for learning to upsample. The upsampled maps are sparse and are then convolved with trainable filters to produce dense feature maps. We compare our proposed architecture with the widely adopted FCN [2] and also with the well known DeepLab-LargeFOV [3] , DeconvNet [4] architectures. This comparison reveals the memory versus accuracy trade-off involved in achieving good segmentation performance. SegNet was primarily motivated by scene understanding applications. Hence, it is designed to be efficient both in terms of memory and computational time during inference. It is also significantly smaller in the number of trainable parameters than other competing architectures and can be trained end-to-end using stochastic gradient descent. We also performed a controlled benchmark of SegNet and other architectures on both road scenes and SUN RGB-D indoor scene segmentation tasks. These quantitative assessments show that SegNet provides good performance with competitive inference time and most efficient inference memory-wise as compared to other architectures. We also provide a Caffe implementation of SegNet and a web demo at http://mi.eng.cam.ac.uk/projects/segnet/.
1. The bordism groups. This note presents an outline of the authors' efforts to apply Thorn's cobordism theory [ó] to the study of differentiable periodic maps. First, however, we shall outline our scheme for computing the oriented bordism groups of a space [ l ] . These preliminary remarks bear on a problem raised by Milnor [4]. A finite manifold is the finite disjoint union of compact connected manifolds with boundary each of which carries a O-differential structure. The boundary of a finite manifold, B, is denoted by dB. A closed manifold is a finite manifold with void boundary. We now define the oriented bordism groups of a pair (X, ^4). An oriented singular manifold in (X, A) is a map ƒ: (B} dB ) —»(X, A) of an oriented finite manifold. Such a singular manifold bords in (X, A) if and only if there is a finite oriented manifold W and a map F: W—->X such that BC.dW as a finite regular submanifold whose orientation is induced by that of W and such that F\ jB=/, F(dW— B) C.A. From two such oriented singular manifolds (Bl fx) and (£?, /2) a disjoint union (B\\JB n 2l fxKJf2) is formed with B\C\B% = 0 and / i U / 2 | £?==ƒ,, * = 1 , 2. Obviously ( £ » , ƒ ) = ( J 3 n , ƒ). We £ay that two singular manifold (5J, /i) and (J5J,/2) are bordant in (X, yl) if and only if the disjoint union (JB*U -~B1,f\\Jf(X, ^4) and any closed oriented manifold V the module product is given by [B, / ] [ F W ] = [BX V, g] where g(x9 y) =ƒ(*). For any map : (X, A)-*(Y, B) there is an induced homomorphism f]. There is also d*: Qn(X, A)-*Qn-i(A) given by d*([5», ƒ ] ) = [3B», f\dB-*A]. Actually 0*: &*(X, i4)-*Q*(F, 5 ) and d*: J2*(X, ^4)~>fts|c(^4) are fl-module homomorphisms of degree 0 and 1 .
Ramin Mousa, Saeed Chamani, Mohammad Morsali
et al.
Skin cancer (SC) is a life-threatening disease where early diagnosis is critical for effective treatment and survival. While deep learning (DL) has advanced skin cancer diagnosis (SCD), current methods generally yield suboptimal accuracy and efficiency due to challenges in extracting multiscale features from dermoscopic images and optimizing complex model parameters through efficient exploration of the space of hyperparameters. To address this, we propose an approach integrating late Discrete Wavelet Transform (DWT) with pre-trained convolutional neural networks (CNNs) and swarm-based optimization. The late DWT decomposes CNN-extracted feature maps into low- and high-frequency components to improve the detection of subtle lesion patterns, while a self-attention mechanism further refines this by weighing feature importance, focusing on relevant diagnostic information. To refine hyperparameters, three novel swarm-based optimizers – Modified Gorilla Troops Optimizer (MGTO), Improved Gray Wolf Optimization (IGWO), and Fox Optimization (FOX) – are employed searching the space of the hyperparameters to fine-tune the model for superior performance. In comparison to existing methods, experiments on the ISIC-2016 and ISIC-2017 datasets show enhanced classification performance, obtaining at least a 1% accuracy gain. Thus, the suggested framework offers a reliable and effective way to diagnose skin cancer automatically.
Objectives
To develop a resource that maps health outcomes across coding schemas in linked administrative data in UK Biobank, addressing the challenge of identifying equivalent outcomes from multiple sources. Our approach minimised the loss of clinical detail, a common limitation in such efforts, to enhance its utility for health research.
Methods
UK Biobank is a prospective cohort study of ~500,000 adults, recruited between 2006-10, with follow up for health outcomes through linkage with administrative health data. Clinical coding schemas include Read Version 2 (Read2) and Clinical Terms Version 3 (CTV3) from primary care, and International Classification of Diseases (ICD) 9th and 10th editions (ICD-9 and ICD-10) from secondary care, cancer registries and death records; self-reported conditions were also reported at recruitment. We reviewed existing mapping resources and, with clinical support, mapped clinical codes in different schemas to 4-digit ICD-10 to provide detailed clinical information using a single internationally-recognised schema.
Results
We processed data from 230,096 participants with primary care records, 442,267 with secondary care records, 40,447 with death records, and 397,063 with self-reported data. We successfully mapped to 81% of Read2 codes (N = 12,448), 93% of CTV3 (24,188), 92% of ICD-9 (3,060), and 100% of self-reported (509) to ICD-10 codes. Although existing resources frequently allowed a single code to be mapped to a single ICD-10 code (94% of the mapped codes for Read2, 58% of CTV3, and 79% of ICD-9), the remaining codes require extensive clinical review, which is ongoing. The conversion increased the granularity of health outcomes by 5.8 times from 2,006 3-digit ICD-10 codes to 11,625 4-digit ICD-10 codes. The most common ICD-10 codes included those related to musculoskeletal diseases (24%).
Conclusion
The increased granularity of ICD coding enhances the research potential of UK Biobank data, enabling precise outcome definitions and detailed comparisons with other healthcare datasets. The enhanced mappings revealed underrepresented and nuanced outcomes, improving subtyping of conditions, and supporting robust comparisons with external datasets using internationally recognised coding standards.
Bouazza Kacimi, Ahmed Mohammed Cherif, Mustafa Özkan
Let $ψ: (M,g)\longrightarrow (N,h)$ be a smooth map between Riemannian manifolds. The tension field of $ψ$ can be regarded as a map from $(M,g)$ into the Riemannian vector bundle $ψ^{-1}TN$, equipped with the Sasaki metric $G_{S}$. In this paper, we study certain aspects of two types of maps: those whose tension fields are harmonic maps (called $HM$-tensional maps) and those whose tension fields are harmonic sections (called $HS$-tensional maps).
Crowd counting is an important task that serves as a preprocessing step in many applications. Despite obvious improvement reported by various convolutional-neural-network-based approaches, they only focus on the role of deep feature maps while neglecting the importance of shallow features for crowd counting. In order to surmount this issue, a dilated convolutional-neural-network-based cross-level contextual information extraction network is proposed in this work, which is abbreviated as CL-DCNN. Specifically, a dilated contextual module (DCM) is constructed by importing cross-level connection between different feature maps. It can effectively integrate contextual information while conserving the local details of crowd scenes. Extensive experiments show that the proposed approach outperforms state-of-the-art approaches using five public datasets, i.e., ShanghaiTech part A, ShanghaiTech part B, Mall, UCF_CC_50 and UCF-QNRF, achieving MAE 52.6, 8.1, 1.55, 181.8, and 96.4, respectively.
Inundation maps that show water depths that occur in the event of a flood are essential for protection. Especially information on timings is crucial. Creating a dynamic inundation map with depth data in temporal resolution is a major challenge and is not possible with physical models, as these are too slow for real-time predictions. To provide a dynamic inundation map in real-time, we developed a data-driven multi-step inundation forecast system for fluvial flood events. The forecast system is based on a convolutional neural network (CNN), feature-informed dense layers, and a recursive connection from the predicted inundation at timestep t as a new input for timestep t + 1. The forecast system takes a hydrograph as input, cuts it at desired timesteps (t), and outputs the respective inundation for each timestep, concluding in a dynamic inundation map with a temporal resolution (t). The prediction shows a Critical Success Index (CSI) of over 90%, an average Root Mean Square Error (RMSE) of 0.07, 0.12, and 0.15 for the next 6 h, 12 h, and 24 h, respectively, and an individual RMSE value below 0.3 m, for all test datasets when compared with the results from a physically based model.
The advancement in technology has made interdisciplinary research more accessible. Particularly the breakthrough in Artificial Intelligence AI has given huge advantages to researchers working in interdisciplinary and multidisciplinary fields. This study investigates the ability of AI models, particularly GPT4 and GPT Data Analyst in creating language maps for language documentation. The study Integrates documentary linguistics linguistic geography and AI by showcasing how AI models facilitate the spatial documentation of languages through the creation of language maps with minimal cartographic expertise. The study is conducted using a CSV file and a GeoJSON file both obtained from HDX and from the researchers fieldwork. The study data is then applied in realtime conversations with the AI models in order to generate the language distribution maps. The study highlights the two AI models capabilities in generating highquality static and interactive web maps and streamlining the mapmaking process, despite facing challenges like inconsistencies and difficulties in adding legends. The findings suggest a promising future for AI in generating language maps and enhancing the work of documentary linguists as they collect their data in the field pointing towards the need for further development to fully harness AI potential in this field.
We consider proper holomorphic maps of ball complements and differences in complex euclidean spaces of dimension at least two. Such maps are always rational, which naturally leads to a related problem of classifying rational maps taking concentric spheres to concentric spheres, what we call $m$-fold sphere maps; a proper map of the difference of concentric balls is a $2$-fold sphere map. We prove that proper maps of ball complements are in one to one correspondence with polynomial proper maps of balls taking infinity to infinity. We show that rational $m$-fold sphere maps of degree less than $m$ (or polynomial maps of degree $m$ or less) must take all concentric spheres to concentric spheres and we provide a complete classification of them. We prove that these degree bounds are sharp.
Let $\mathsf{Q}$ be a commutative and unital quantale. By a $\mathsf{Q}$-map we mean a left adjoint in the quantaloid of sets and $\mathsf{Q}$-relations, and by a partial $\mathsf{Q}$-map we refer to a Kleisli morphism with respect to the maybe monad on the category $\mathsf{Q}\text{-}\mathbf{Map}$ of sets and $\mathsf{Q}$-maps. It is shown that every $\mathsf{Q}$-map is symmetric if and only if $\mathsf{Q}$ is weakly lean, and that every $\mathsf{Q}$-map is exactly a map in $\mathbf{Set}$ if and only $\mathsf{Q}$ is lean. Moreover, assuming the axiom of choice, it is shown that the category of sets and partial $\mathsf{Q}$-maps is monadic over $\mathsf{Q}\text{-}\mathbf{Map}$.
Robel T Beyene, David P Stonko, Stephen P Gondek
et al.
<h4>Background</h4>Variations in admission patterns have been previously identified in non-elective surgical services, but minimal data on the subject exists with respect to burn admissions. Improved understanding of the temporal pattern of burn admissions could inform resource utilization and clinical staffing. We hypothesize that burn admissions have a predictable temporal distribution with regard to the time of day, day of week, and season of year in which they present.<h4>Study design</h4>A retrospective, cohort observational study of a single burn center from 7/1/2016 to 3/31/2021 was performed on all admissions to the burn surgery service. Demographics, burn characteristics, and temporal data of burn admissions were collected. Bivariate absolute and relative frequency data was captured and plotted for all patients who met inclusion criteria. Heat-maps were created to visually represent the relative admission frequency by time of day and day of week. Frequency analysis grouped by total body surface area against time of day and relative encounters against day of year was performed.<h4>Results</h4>2213 burn patient encounters were analyzed, averaging 1.28 burns per day. The nadir of burn admissions was from 07:00 and 08:00, with progressive increase in the rate of admissions over the day. Admissions peaked in the 15:00 hour and then plateaued until midnight (p<0.001). There was no association between day of week in the burn admission distribution (p>0.05), though weekend admissions skewed slightly later (p = 0.025). No annual, cyclical trend in burn admissions was identified, suggesting that there is no predictable seasonality to burn admissions, though individual holidays were not assessed.<h4>Conclusion</h4>Temporal variations in burn admissions exist, including a peak admission window late in the day. Furthermore, we did not find a predictable annual pattern to use in guiding staffing and resource allocation. This differs from findings in trauma, which identified admission peaks on the weekends and an annual cycle that peaks in spring and summer.