Custom Loss Function Research Articles

Temporal attention fusion network with custom loss function for EEG--fNIRS classification.

Objective.Methods that can detect brain activities accurately are crucial owing to the increasing prevalence of neurological disorders. In this context, a combination of electroencephalography (EEG) and functional near-infrared spectroscopy (fNIRS) offers a powerful approach to understanding normal and pathological brain functions, thereby overcoming the limitations of each modality, such as susceptibility to artefacts of EEG and limited temporal resolution of fNIRS. However, challenges such as class imbalance and inter-class variability within multisubject data hinder their full potential.Approach.To address this issue, we propose a novel temporal attention fusion network (TAFN) with a custom loss function. The TAFN model incorporates attention mechanisms to its long short-term memory and temporal convolutional layers to accurately capture spatial and temporal dependencies in the EEG--fNIRS data. The custom loss function combines class weights and asymmetric loss terms to ensure the precise classification of cognitive and motor intentions, along with addressing class imbalance issues.Main resultsRigorous testing demonstrated the exceptional cross-subject accuracy of the TAFN, exceeding 99% for cognitive tasks and 97% for motor imagery (MI) tasks. Additionally, the ability of the model to detect subtle differences in epilepsy was analyzed using scalp topography in MI tasks.Significance.This study presents a technique that outperforms traditional methods for detecting high-precision brain activity with subtle differences in the associated patterns. This makes it a promising tool for applications such as epilepsy and seizure detection, in which discerning subtle pattern differences is of paramount importance.

Sales Forecasting with LSTM, Custom Loss Function, and Hyperparameter Optimization: A Case Study

Forecasting sales trends is a valuable activity for companies of all types and sizes, as it enables more efficient decision making to avoid unnecessary expenses from excess inventory or, conversely, losses due to insufficient inventory to meet demand. In this paper, we designed a personalized cost function to reduce economic losses caused by the excessive acquisition of products or derived from their scarcity when needed. Moreover, we designed an LSTM network integrated with Glorot and Orthogonal initializers and dropout to forecast sales trends in a lumber mill in Tamaulipas, Mexico. To generalize and appropriately forecast the sales of the lumber mill products, we optimized the LSTM network’s hyperparameters through a genetic algorithm, which was essential to explore the solution space. We evaluated our proposal in instances obtained from the historical sales of the five main products sold by the lumber mill. According to the results, we concluded that for our case study the proposed function cost and the hyperparameters optimization allowed the LSTM to forecast the direction and trend of the lumber mill’s product sales despite the variability of the products.

Artificial intelligence-guided screening of under-recognized cardiomyopathies adapted for point-of-care echocardiography

Abstract Background Point-of-care ultrasonography (POCUS) enables access to cardiac imaging directly at the bedside but is limited by brief acquisition, variation in acquisition quality, and lack of advanced protocols. Purpose To develop and validate deep learning models for detecting underdiagnosed cardiomyopathies on cardiac POCUS, leveraging a novel acquisition quality-adapted modeling strategy. Methods To develop the models, we identified transthoracic echocardiograms (TTEs) of patients across five hospitals in a large U.S. health system with transthyretin amyloid cardiomyopathy (ATTR-CM, confirmed by Tc99m-pyrophosphate imaging), hypertrophic cardiomyopathy (HCM, confirmed by cardiac magnetic resonance), and random controls from the same period enriched for the presence of severe AS. In a sample of 290,245 TTE videos, we used novel augmentation approaches and a customized loss function to weigh image and view quality to train a multi-label, view agnostic video-based convolutional neural network (CNN) to discriminate the presence of ATTR-CM, HCM, and/or AS+ and AS- controls. Models were tested across 3,758 real-world POCUS videos from 1,879 studies in 1,330 independent emergency department (ED) patients from 2011 through 2023 (A). Results Our multi-label, view-agnostic classifier demonstrated state-of-the-art performance in discriminating ATTR-CM (AUROC 0.98 [95%CI: 0.96-0.99]) and HCM (AUROC 0.95 [95% CI: 0.94-0.96]) on standard TTE studies. Automated metrics of anatomical view correctness confirmed significantly lower quality in POCUS vs TTE videos (median view classifier confidence of 0.63 [IQR: 0.44-0.88] vs 0.93 [IQR: 0.69-1.00], p<0.001). When deployed to POCUS videos, our algorithm effectively discriminated ATTR-CM and HCM with AUROC of up to 0.94 (parasternal long-axis (PLAX)), and 0.85 (apical 4 chamber (A4C)) (B), corresponding to positive diagnostic odds ratios of 46.7, and 25.5, respectively. In total, 18/35 (51.4%) of ATTR-CM and 32/57 (41.1%) of HCM patients in the POCUS cohort had a positive AI screen in the year before their eventual confirmatory imaging (C). Conclusions We define and validate an AI framework that enables scalable, opportunistic screening of under-diagnosed cardiomyopathies using POCUS.

Inspection robot GPS outages localization based on error Kalman filter and deep learning

In urban environments, inspection robots face complex terrain and variable motion states, posing high demands on their positioning systems. Although the integration of Micro-Electro-Mechanical Systems Inertial Navigation Systems (MEMS-INS) with the Global Positioning System (GPS) provides continuous positioning information, high buildings and tunnels in cities can block GPS signals, leading to signal interruptions and increased positioning errors. During GPS outages, MEMS-INS gradually accumulates errors, severely affecting positioning accuracy. To address this issue, this paper proposes an adaptive error state Kalman Filter (AESKF), which employs an adaptive mechanism to eliminate the noise impact of MEMS-INS and reduce reliance on the process model. Additionally, a deep learning framework based on the Self-Attention mechanism of the Transformer and a custom loss function Long Short-Term Memory (LSTM) module is proposed to predict position increments of the inspection robot. Combining AESKF with Transformer-LSTM achieves optimized positioning accuracy of the inspection robot during GPS outages in dynamic urban environments. Simulation and practical experimental results demonstrate that the combination of AESKF and Transformer-LSTM significantly improves positioning accuracy. Compared to other mature methods, the Root Mean Square Error (RMSE) of positioning is reduced by up to 83.64 % in the north direction and 89.56 % in the east direction. When the GPS signal interruption lasts for 10 s and 60 s, the maximum position error standard deviation (STD) is 0.1186 m and 1.0417 m, respectively.

Hydro-informer: a deep learning model for accurate water level and flood predictions

AbstractThis study aims to develop an advanced deep learning model, Hydro-Informer, for accurate water level and flood predictions, emphasizing extreme event forecasting. Utilizing a comprehensive dataset from the Slovak Hydrometeorological Institute SHMI (2008–2020), which includes precipitation, water level, and discharge data, the model was trained using a ladder technique with a custom loss function to enhance focus on extreme values. The architecture integrates Recurrent and Convolutional Neural Networks (RNN, CNN), and Multi-Head Attention layers. Hydro-Informer achieved significant performance, with a Coefficient of Determination (R2) of 0.88, effectively predicting extreme water levels 12 h in advance in a river environment free from human regulation and structures. The model’s strong performance in identifying extreme events highlights its potential for enhancing flood management and disaster preparedness. By integrating with diverse data sources, the model can be used to develop a well-functioning warning system to mitigate flood impacts. This work proposes a novel architecture suitable for locations without water regulation structures.

Efficient property-oriented design of composite layups via controllable latent features using generative VAE

Fiber-reinforced composites provide substantial tailoring potential, while the extensive parameters and complex coupling mechanisms pose formidable challenges to layup designs. This paper presents an efficient inverse design framework for composite layups utilizing a variational autoencoder (VAE), which is applicable to non-conventional laminates. By leveraging the VAE's exceptional feature extraction and generative capabilities, the decoder rapidly produces layups with desired properties through controllable feature vectors. Based on the stacking characteristics of layups, multi-scale one-dimensional convolutions precisely extract sequence features relevant to mechanical properties and specific manufacturing constraints. A customized loss function is formulated to constrain the latent features, while addressing the non-uniqueness problem for layups with certain mechanical properties. The developed property-oriented VAE can generate 100,000 layups in seconds, achieving an average success rate of 66.9% under comprehensive in-plane and bending stiffness design, and remains effective for 100-ply thick laminate. For comparison, the VAE model outperforms the genetic algorithm and the logic-based method in reinforced panel designs, reducing the retrieval error by 46.4% and 38.1%, respectively. The proposed approach demonstrates flexible and efficient design advantages using generative machine learning models, and is easily extendable to other inverse design scenarios.

In-situ monitoring method of femtosecond laser welding between glass and copper with acoustic emission

The use of femtosecond laser welding has created new possibilities for connecting glasses and metals, revolutionizing the fabrication and assembly of advanced electronic components. However, monitoring the connection process during femtosecond laser welding is challenging due to the small heat-affected zone and brief interaction time. This study utilizes acoustic emission to monitor femtosecond laser welding and introduces the “Sum of Relative RMS” feature to correlate connection formation with acoustic emission signals in glass-copper welding. The mechanism and signal occurrence trends during welding are explained through in-situ signal monitoring and offline scanning electron microscopy characterization. Additionally, a Convolutional Neural Network is enhanced by increasing the convolutional kernel size, integrating multi-dimensional features, and incorporating prior knowledge via a customized loss function, resulting in an increase in the accuracy of successful welding detection from 89% to 95%. This research provides new insights into the monitoring of femtosecond laser glass-metal heterogeneous welding.

Deep Learning Regression with Sequences of Different Length: An Application for State of Health Trajectory Prediction and Remaining Useful Life Estimation in Lithium-Ion Batteries

This study presents methods to handle deep learning regressions with input and output sequences of different lengths. We discuss the Autoregressive one-step prediction framework and introduce an innovative one-time multi-step (OTMS) prediction approach, based on a custom loss function, that predicts all future steps in a single shot. The presented methodologies are then applied to simultaneously predict the State of Health (SoH) trajectory and estimate the Remaining Useful Life (RUL) of lithium-ion battery cells. Accurate estimates of SoH trajectory and RUL are essential for Battery Management Systems (BMSs), electronic systems that guarantee safety while maximizing performance and extending battery lifespan. In this context, the studied methodologies were compared using a rigorous cross-validation approach. The OTMS model showed better predictions in early cycles, while the Autoregressive model performed better in later cycles, suggesting a hybrid approach between these two methodologies as an optimal solution.

Designing observables for measurements with deep learning

Many analyses in particle and nuclear physics use simulations to infer fundamental, effective, or phenomenological parameters of the underlying physics models. When the inference is performed with unfolded cross sections, the observables are designed using physics intuition and heuristics. We propose to design targeted observables with machine learning. Unfolded, differential cross sections in a neural network output contain the most information about parameters of interest and can be well-measured by construction. The networks are trained using a custom loss function that rewards outputs that are sensitive to the parameter(s) of interest while simultaneously penalizing outputs that are different between particle-level and detector-level (to minimize detector distortions). We demonstrate this idea in simulation using two physics models for inclusive measurements in deep inelastic scattering. We find that the new approach is more sensitive than classical observables at distinguishing the two models and also has a reduced unfolding uncertainty due to the reduced detector distortions.

Deep learning-based voxel sampling for particle therapy treatment planning

Objective. Scanned particle therapy often requires complex treatment plans, robust optimization, as well as treatment adaptation. Plan optimization is especially complicated for heavy ions due to the variable relative biological effectiveness. We present a novel deep-learning model to select a subset of voxels in the planning process thus reducing the planning problem size for improved computational efficiency. Approach. Using only a subset of the voxels in target and organs at risk (OARs) we produced high-quality treatment plans, but heuristic selection strategies require manual input. We designed a deep-learning model based on P-Net to obtain an optimal voxel sampling without relying on patient-specific user input. A cohort of 70 head and neck patients that received carbon ion therapy was used for model training (50), validation (10) and testing (10). For training, a total of 12 500 carbon ion plans were optimized, using a highly efficient artificial intelligence (AI) infrastructure implemented into a research treatment planning platform. A custom loss function increased sampling density in underdosed regions, while aiming to reduce the total number of voxels. Main results. On the test dataset, the number of voxels in the optimization could be reduced by 84.8% (median) at <1% median loss in plan quality. When the model was trained to reduce sampling in the target only while keeping all voxels in OARs, a median reduction up to 71.6% was achieved, with 0.5% loss in the plan quality. The optimization time was reduced by a factor of 7.5 for the total AI selection model and a factor of 3.7 for the model with only target selection. Significance. The novel deep-learning voxel sampling technique achieves a significant reduction in computational time with a negligible loss in the plan quality. The reduction in optimization time can be especially useful for future real-time adaptation strategies.

Multiscale segmentation net for segregating heterogeneous brain tumors: Gliomas on multimodal MR images

In this research, the 3D volumetric segmentation of heterogeneous brain tumors such as Gliomas- anaplastic astrocytoma, and Glioblastoma Multiforme (GBM) is performed to extract enhancing tumor (ET), whole tumor (WT), and tumor core (TC) regions using T1, T2, and FLAIR images. Therefore, a deep learning-based encoder-decoder architecture named “MS-SegNet” using 3D multi-scale convolutional layers is proposed. The proposed architecture employs multi-scale feature extraction (MS-FE) block the filter size 3 × 3 × 3 to extract confined information like tumor boundary and edges of necrotic part. The filter of size 5 × 5 × 5 focuses on varied features like shape, size, and location of tumor region with edema. The local and global features from different MR modalities are extracted for segmenting thin and meshed boundaries of heterogeneous tumors between anatomical sub-regions like peritumoral edema, enhancing tumor, and necrotic tumor core. The learning parameters on introducing the MS-FE block are reduced to 10 million, which is much less than other architectures like 3D-Unet which takes into consideration 27 million features leading to the consumption of less computational power. A customized loss function is also prophesied based on a combination of dice loss and focal loss along with metrics such as accuracy and Intersection over Union (IoU) i.e. the overlapping of ground truth mask and predicted value for addressing the class imbalance problem. For evaluating the efficacy of the proposed method, four evaluation metrics such as Dice Coefficient (DSC), Sensitivity, Specificity, and Hausdorff95 distance (H95) are employed for analyzing the model's overall performance. It is observed that the proposed MS-SegNet architecture achieved the DSC of 0.81, 0.91, and 0.83 on BraTS 2020; 0.86, 0.92, and 0.84 on BraTS 2021 for ET, WT, and TC respectively. The developed model is also tested on a real-time dataset collected from the Post Graduate Institute of Medical Education & Research (PGIMER), Chandigarh. The DSC of 0.79, 0.76, and 0.68 for ET, WT, and TC respectively on the real-time dataset. These findings show that deep learning models with enhanced feature extraction capabilities can be readily trained to attain high accuracy in segmenting heterogeneous brain tumors and hold promising results. In the future, other tumor datasets will be explored for the detection and treatment planning of brain tumors to check the effectiveness of the model in real-world healthcare environments.

Artificial intelligence-guided detection of under-recognized cardiomyopathies on point-of-care cardiac ultrasound: a multi-center study.

Point-of-care ultrasonography (POCUS) enables cardiac imaging at the bedside and in communities but is limited by abbreviated protocols and variation in quality. We developed and tested artificial intelligence (AI) models to automate the detection of underdiagnosed cardiomyopathies from cardiac POCUS. In a development set of 290,245 transthoracic echocardiographic videos across the Yale-New Haven Health System (YNHHS), we used augmentation approaches and a customized loss function weighted for view quality to derive a POCUS-adapted, multi-label, video-based convolutional neural network (CNN) that discriminates HCM (hypertrophic cardiomyopathy) and ATTR-CM (transthyretin amyloid cardiomyopathy) from controls without known disease. We evaluated the final model across independent, internal and external, retrospective cohorts of individuals who underwent cardiac POCUS across YNHHS and Mount Sinai Health System (MSHS) emergency departments (EDs) (2011-2024) to prioritize key views and validate the diagnostic and prognostic performance of single-view screening protocols. We identified 33,127 patients (median age 61 [IQR: 45-75] years, n=17,276 [52·2%] female) at YNHHS and 5,624 (57 [IQR: 39-71] years, n=1,953 [34·7%] female) at MSHS with 78,054 and 13,796 eligible cardiac POCUS videos, respectively. An AI-enabled single-view screening approach successfully discriminated HCM (AUROC of 0·90 [YNHHS] & 0·89 [MSHS]) and ATTR-CM (YNHHS: AUROC of 0·92 [YNHHS] & 0·99 [MSHS]). In YNHHS, 40 (58·0%) HCM and 23 (47·9%) ATTR-CM cases had a positive screen at median of 2·1 [IQR: 0·9-4·5] and 1·9 [IQR: 1·0-3·4] years before clinical diagnosis. Moreover, among 24,448 participants without known cardiomyopathy followed over 2·2 [IQR: 1·1-5·8] years, AI-POCUS probabilities in the highest (vs lowest) quintile for HCM and ATTR-CM conferred a 15% (adj.HR 1·15 [95%CI: 1·02-1·29]) and 39% (adj.HR 1·39 [95%CI: 1·22-1·59]) higher age- and sex-adjusted mortality risk, respectively. We developed and validated an AI framework that enables scalable, opportunistic screening of treatable cardiomyopathies wherever POCUS is used. National Heart, Lung and Blood Institute, Doris Duke Charitable Foundation, BridgeBio.

Removing ring artifacts in CBCT images via Transformer with unidirectional vertical gradient loss.

Cone-beam computed tomography (CBCT), an important medical modality for disease detection and diagnosis, is currently widely used in clinical practice. However, due to the inconsistent response of CBCT detectors, the lack of proper calibration often leads to the occurrence of ring artifacts in CBCT-reconstructed images. These artifacts may affect physicians' assessment and diagnosis. Therefore, effective elimination of ring artifacts in CBCT images without degrading image quality isimportant. Given the pros and cons of existing methods for removing ring artifacts in CBCT images, this paper is devoted to designing a specific Transformer for this task, leveraging the global and local modeling ability ofTransformer. We design a loss function with dual-domain information fusion for the vanilla Transformer to correct ring artifacts in CBCT images. The method operates in image domain to predict artifact-free outputs and preserve more image details. Meanwhile, we design a tailored loss function incorporating polar domain optimization to remove ring artifacts more effectively. Specifically, an unidirectional gradient loss that constrains vertical gradients in polar domain is imposed, based on the geometric prior that in polar coordinates, ring artifacts predominately affect horizontal gradients while minimally influencing verticalgradients. We conduct extensive ablative and comparative experiments on CBCT/CT image set to validate the performance of the proposed method. First, four ablation experiments demonstrate the feasibility of our approach. Then, we compare our method with several classical methods and the latest state-of-the-arts, and our method achieves the highest quality of corrected images as well as the best evaluation metrics. In these experiments, 5332 CT images were used for training, and 550 CT images, and 500 real CBCT images were used for testing. The source code is available at https://github.com/shasha521/CBCT. Experimental results demonstrate that our method can significantly improve the effectiveness of ring artifact correction. By capitalizing on dual-domain information fusion and a customized loss function, the improved Transformer can not only effectively remove ring artifacts in CBCT images, but also preserve the details of original images quitewell.

Patent lifetime prediction using LightGBM with a customized loss

Patent lifespan is commonly used as a quantitative measure in patent assessments. Patent holders maintain exclusive rights by paying significant maintenance fees, suggesting a strong correlation between a patent’s lifespan and its business potential or economic value. Therefore, accurately forecasting the duration of a patent is of great significance. This study introduces a highly effective method that combines LightGBM, a sophisticated machine learning algorithm, with a customized loss function derived from Focal Loss. The purpose of this approach is to accurately predict the probability of a patent remaining valid until its maximum expiration date. This research differs from previous studies that have examined the various stages and phases of patents. Instead, it assesses the commercial viability of individual patents by considering their lifespan. The evaluation process utilizes a dataset consisting of 200,000 patents. The experimental results show a significant improvement in the performance of the model by combining Focal Loss with LightGBM. By incorporating Focal Loss into LightGBM, its ability to give priority to difficult instances during training is enhanced, resulting in an overall improvement in performance. This targeted approach enhances the model’s ability to distinguish between different samples and its ability to recover from challenges by giving priority to difficult samples. As a result, it improves the model’s accuracy in making predictions and its ability to apply those predictions to new data.

Transforming Poultry Farming: A Pyramid Vision Transformer Approach for Accurate Chicken Counting in Smart Farm Environments.

Smart farm environments, equipped with cutting-edge technology, require proficient techniques for managing poultry. This research investigates automated chicken counting, an essential part of optimizing livestock conditions. By integrating artificial intelligence and computer vision, it introduces a transformer-based chicken-counting model to overcome challenges to precise counting, such as lighting changes, occlusions, cluttered backgrounds, continual chicken growth, and camera distortions. The model includes a pyramid vision transformer backbone and a multi-scale regression head to predict precise density maps of the crowded chicken enclosure. The customized loss function incorporates curriculum loss, allowing the model to learn progressively, and adapts to diverse challenges posed by varying densities, scales, and appearances. The proposed annotated dataset includes data on various lighting conditions, chicken sizes, densities, and placements. Augmentation strategies enhanced the dataset with brightness, contrast, shadow, blur, occlusion, cropping, and scaling variations. Evaluating the model on the proposed dataset indicated its robustness, with a validation mean absolute error of 27.8, a root mean squared error of 40.9, and a test average accuracy of 96.9%. A comparison with the few-shot object counting model SAFECount demonstrated the model's superior accuracy and resilience. The transformer-based approach was 7.7% more accurate than SAFECount. It demonstrated robustness in response to different challenges that may affect counting and offered a comprehensive and effective solution for automated chicken counting in smart farm environments.

Leveraging a realistic synthetic database to learn Shape-from-Shading for estimating the colon depth in colonoscopy images

Colonoscopy is the choice procedure to diagnose, screening, and treat the colon and rectum cancer, from early detection of small precancerous lesions (polyps), to confirmation of malign masses. However, the high variability of the organ appearance and the complex shape of both the colon wall and structures of interest make this exploration difficult. Learned visuospatial and perceptual abilities mitigate technical limitations in clinical practice by proper estimation of the intestinal depth. This work introduces a novel methodology to estimate colon depth maps in single frames from monocular colonoscopy videos. The generated depth map is inferred from the shading variation of the colon wall with respect to the light source, as learned from a realistic synthetic database. Briefly, a classic convolutional neural network architecture is trained from scratch to estimate the depth map, improving sharp depth estimations in haustral folds and polyps by a custom loss function that minimizes the estimation error in edges and curvatures. The network was trained by a custom synthetic colonoscopy database herein constructed and released, composed of 248400 frames (47 videos), with depth annotations at the level of pixels. This collection comprehends 5 subsets of videos with progressively higher levels of visual complexity. Evaluation of the depth estimation with the synthetic database reached a threshold accuracy of 95.65%, and a mean-RMSE of 0.451cm, while a qualitative assessment with a real database showed consistent depth estimations, visually evaluated by the expert gastroenterologist coauthoring this paper. Finally, the method achieved competitive performance with respect to another state-of-the-art method using a public synthetic database and comparable results in a set of images with other five state-of-the-art methods. Additionally, three-dimensional reconstructions demonstrated useful approximations of the gastrointestinal tract geometry. Code for reproducing the reported results and the dataset are available at https://github.com/Cimalab-unal/ColonDepthEstimation.

Turbulence modeling of stratified turbulence using a constrained artificial neural network

For large eddy simulations (LES) of stratified turbulence in the strongly stratified regime, an artificial neural network (ANN) with five hidden layers is used to construct a sub-grid scale (SGS) model. The ANN is assessed by comparing it to the Smagorinsky model, the dynamic Smagorinsky model, the gradient model, and filtered direct numerical simulation data. In the a priori test, the SGS model using ANN performed better than the Smagorinsky model and the gradient model in terms of the correlation coefficient and relative error of the energy transfer rate. However, the ANN does not provide sufficient energy dissipation when it is applied to LES with a larger filter width because it overpredicts backscatter. To address this problem, we also trained a constrained ANN using a custom loss function that penalizes excessive backscatter. It is shown that the constrained ANN successfully predicts less backscatter, maintaining the high correlation coefficient without ad hoc clipping. These results show that ANN is a promising tool for realizing a highly accurate and stable SGS model for stratified turbulence.

LUNet: deep learning for the segmentation of arterioles and venules in high resolution fundus images

Objective. This study aims to automate the segmentation of retinal arterioles and venules (A/V) from digital fundus images (DFI), as changes in the spatial distribution of retinal microvasculature are indicative of cardiovascular diseases, positioning the eyes as windows to cardiovascular health. Approach. We utilized active learning to create a new DFI dataset with 240 crowd-sourced manual A/V segmentations performed by 15 medical students and reviewed by an ophthalmologist. We then developed LUNet, a novel deep learning architecture optimized for high-resolution A/V segmentation. The LUNet model features a double dilated convolutional block to widen the receptive field and reduce parameter count, alongside a high-resolution tail to refine segmentation details. A custom loss function was designed to prioritize the continuity of blood vessel segmentation. Main Results. LUNet significantly outperformed three benchmark A/V segmentation algorithms both on a local test set and on four external test sets that simulated variations in ethnicity, comorbidities and annotators. Significance. The release of the new datasets and the LUNet model (www.aimlab-technion.com/lirot-ai) provides a valuable resource for the advancement of retinal microvasculature analysis. The improvements in A/V segmentation accuracy highlight LUNet's potential as a robust tool for diagnosing and understanding cardiovascular diseases through retinal imaging.

Neural-network based high-speed volumetric dynamic optical coherence tomography

We demonstrate deep-learning neural network (NN)-based dynamic optical coherence tomography (DOCT), which generates high-quality logarithmic-intensity-variance (LIV) DOCT images from only four OCT frames. The NN model is trained for tumor spheroid samples using a customized loss function: the weighted mean absolute error. This loss function enables highly accurate LIV image generation. The fidelity of the generated LIV images to the ground truth LIV images generated using 32 OCT frames is examined via subjective image observation and statistical analysis of image-based metrics. Fast volumetric DOCT imaging with an acquisition time of 6.55 s/volume is demonstrated using this NN-based method.

Mitigating the Impact of Inaccurate Feedback in Dynamic Learning-to-Rank: A Study of Overlooked Interesting Items

Dynamic Learning-to-Rank (DLTR) is a method of updating a ranking policy in real-time based on user feedback, which may not always be accurate. Although previous DLTR work has achieved fair and unbiased DLTR under inaccurate feedback, they face the trade-off between fairness and user utility and also have limitations in the setting of feeding items. Existing DLTR works improve ranking utility by eliminating bias from inaccurate feedback on observed items, but the impact of another pervasive form of inaccurate feedback, overlooked or ignored interesting items, remains unclear. For example, users may browse the rankings too quickly to catch interesting items or miss interesting items because the snippets are not optimized enough. This phenomenon raises two questions: i) Will overlooked interesting items affect the ranking results? ii) Is it possible to improve utility without sacrificing fairness if these effects are eliminated? These questions are particularly relevant for small and medium-sized retailers who are just starting out and may have limited data, leading to the use of inaccurate feedback to update their models. In this paper, we find that inaccurate feedback in the form of overlooked interesting items has a negative impact on DLTR performance in terms of utility. To address this, we treat the overlooked interesting items as noise and propose a novel DLTR method, the Co-teaching Rank (CoTeR), that has good utility and fairness performance when inaccurate feedback is present in the form of overlooked interesting items. Our solution incorporates a co-teaching-based component with a customized loss function and data sampling strategy, as well as a mean pooling strategy to further accommodate newly added products without historical data. Through experiments, we demonstrate that CoTeRx not only enhances utilities but also preserves ranking fairness, and can smoothly handle newly introduced items.