Incremental Training Research Articles

INFLUÊNCIA DOS EXERCÍCIOS DE ALONGAMENTO DA PARTE SUPERIOR DO CORPO NO ARREMESSO DE PESO

ABSTRACT Introduction The technique in the shot put and the ability in the throwing stage are two important factors that determine the ability of athletes. Qualified experts and coaches attach great importance to training and research on the ability of the throwing stage and the ability to throw. Objective Compare gravity load training and single incremental load training through practical means, analyzing the impacts on throwing ability in athletes. Methods The self-assessment method was used to conduct the comparative experiment on different forms of strength training in 20 college students. The experimental scheme adopts the single incremental load strength training, and the control scheme adopts the traditional barbell training. Results : After the experiment, the hand angle of group I was significantly higher than that of group II (P < 0.05); after the experiment, the shoulder angle of group I was significantly higher than that of group II (P < 0.05); after the experiment, the performance of the seated shot placed in group I was significantly higher than that of group II (P < 0.05), and there was no significant difference in the standing performance (P > 0.05). Conclusion Single incremental load training can significantly improve the throwing ability of shot put athletes. Level of evidence II; Therapeutic studies - investigation of treatment outcomes.

The Surprising Effectiveness of Deep Orthogonal Procrustes Alignment in Unsupervised Domain Adaptation

Unsupervised domain adaptation (UDA) aims to transfer and adapt knowledge from a labeled source domain to an unlabeled target domain. Traditionally, geometry-based alignment methods, e.g., Orthogonal Procrustes Alignment (OPA), formed an important class of solutions to this problem. Despite their mathematical tractability, they rarely produce effective adaptation performance with the recent benchmarks. Instead, state-of-the-art approaches rely on sophisticated distribution alignment strategies such as adversarial training. In this paper, we show that, conventional OPA, when coupled with powerful deep feature extractors and a novel bi-level optimization formulation, is indeed an effective choice for handling challenging distribution shifts. When compared to existing UDA methods, our approach offers the following benefits: <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">computational efficiency</i> : Through the isolation of alignment and classifier training steps during adaptation, and the use of deep OPA, our approach is computationally very effective (typically requiring only 700 K parameters more than the base feature extractor as compared to millions of extra parameters required by state-of-the-art UDA baselines); (ii) <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">data efficiency</i> : Our approach does not require updating our feature extractor during adaptation and hence can be effective even with limited target data; (iii) <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">improved generalization</i> : The resulting models are intrinsically well-regularized and demonstrate effective generalization even in the challenging partial DA setting, <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">i.e</i> ., target domain contains only a subset of the classes observed in the source domain.; and (iv) <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">incremental training</i> : Our approach allows progressive adaptation of models to novel domains (unseen during training) without requiring retraining of the model from scratch.

SiFu: A Generic and Robust Multimodal Signal Fusion Platform for Pervasive Localization

We consider pervasive localization for a roaming user who may sample widely varying signal modes (GPS, WiFi, geomagnetism, Bluetooth low energy, etc.) and values over time and space. Previous works can only apply to specific (two or three) modes and environment, and cannot accommodate arbitrary signal modes and environmental changes due to signal noise, device heterogeneity, phone carriage states, etc. We propose SiFu, a simple, accurate, and generic multimodal signal fusion platform robust against environmental deviation from its design point. As a generic framework, SiFu treats a single-modal localization algorithm as a black box to embrace any existing, emerging or future signals with only incremental training. It employs Bayesian deep learning and data augmentation to mitigate the location bias of the single-modal localization algorithm and run-time deviation from the training data, respectively. Using a unified multimodal likelihood formulation and particle filter, it fuses with inertial sensor measurements for localization. We conduct extensive experiments in different venues (campus, mall, and subway station), and show that SiFu achieves significantly higher localization accuracy as compared to state-of-the-art (cutting the error by more than 20%). It is also robust against environmental variations (reducing error by 30%), even when the signal values deviate greatly from their original design settings.

EFEITOS DO TREINAMENTO COM CARGAS PROGRESSIVAS NA RESISTÊNCIA CARDIORRESPIRATÓRIA DE VELOCISTAS

ABSTRACT Introduction: The most evident change during exercise is the increase in oxygen demand, and prolonged high-intensity exercises may impact this even more on this demand. Physical training with progressive loads is a method that aims to reach the training goals by adjusting factors such as initial load, stage load, and a load of the athletes and can be a monitor of the aerobic and anaerobic exercise on the athlete with potential benefits for the high demand of oxygenation. Objective: Study the changes in heart rate and indexes of cardiorespiratory endurance of sprinters under training with progressive loads. Methods: By random sampling, the study subjects were 15 male sprinters and 15 non-sportsmen as research volunteers. The study subjects performed a protocol with progressive load training on a platform. All subjects underwent the Italian Kosmai cardiopulmonary function test before and after training. Results: The subjects who did not practice the sport showed VO2 max values at an exercise intensity of 13 km/h. The VO2 max maximum for sprinters was 15 km/h. Individuals who were not sprinters generally showed lower VT levels than sprinters (P &lt;0.05). Average non-speeder individuals demonstrated higher MET scores than sprinters. Conclusion: Cardiopulmonary endurance and energy metabolism levels are strongly related to long-term participation in physical activity and exercise intensity. Incremental load training improved cardiorespiratory fitness in sprinters. The results may provide a specific theoretical basis for the daily training of sprinters in the future. Level of evidence II; Therapeutic studies - investigation of treatment outcomes.

Text Classification using Incremental Learning for many classes of Support Vector Machines

This study introduces class-incremental learning (CIL). This framework improves multi-class data support vector machines (SVMs). CIL is incremental feature selection and training. Both steps update SVM classifiers as new classes are added to a system. CIL learns one binary sub-classifier and reuses classifier models for subsequent classes. Another feature selection stage follows. Testing applies the classifier to subspacerelevant vector projections. The CIL system is compatible with any binary classification approaches that are used for text analysis. According to the findings of our research, the CIL-based SVM beat the most popular batch SVM learning methods, such as divide-by-2, 1-against-1, and 1-against-rest, and trained faster. These learning methods are used to train SVMs.

ANÁLISE DE HEMOGLOBINA APÓS TREINAMENTO DE SOBRECARGA EM ATLETAS

ABSTRACT Introduction: The cardiovascular system provides athletes with the proper conditions for blood circulation, ensuring the stability and normal metabolism of the body's internal environment during exercise. Objective: Investigate the effect of overload training on the hemoglobin of male taekwondo athletes. Methods: Twenty-one male taekwondo athletes (level 2 or higher) were selected and trained for four weeks, five days per week, with an initial load intensity of 60% of the maximum heart rate and a weekly intensity increase of 10%. Before training and on every weekend during training, hemoglobin (Hb), mean corpuscular volume (MCV), mean corpuscular hemoglobin (MCH), mean corpuscular hemoglobin concentration (MCHC), reticulocyte count (Ret) were checked, RBC volume distribution width (RDW), hemoglobin content distribution width (HDW), mean reticulocyte volume (MCVr), mean reticulocyte hemoglobin concentration (CHCMr), serum iron (Fe) and ferritin (Fer). Results: After four weeks of increasing load training, athletes showed a progressive and significant decrease in Hb (P&lt;0.01), manifested as exercise-induced hypohemoglobin, and MCV, MCH, MCHC, CHCMr, HDW, and serum ferritin were significantly or extremely decreased(P&lt;0.) 05, P&lt;0.01); RDW increased significantly (P&lt;0.05); Changes in Ret and serum iron showed no statistical significance (P&gt;0.05). Correlation analysis found that hemoglobin concentration had the highest correlation with MCHC and CHCMr. Conclusion: Four-week incremental load training can induce exercise-induced hypohemoglobin in male taekwondo athletes, its changes being most correlated with MCHC and CHCMr, but without significant correlation with serum Fe, RDW, HDW, and MCV. Level of Evidence: Therapeutic Studies -Investigation of Outcomes.

Few-Shot Class-Incremental SAR Target Recognition Based on Hierarchical Embedding and Incremental Evolutionary Network

It is difficult to realize effective synthetic aperture radar (SAR) automatic target recognition (ATR) in open scenarios because the ATR model cannot continuously learn from new classes with limited training samples. When adding new classes to the previously trained model, the capability of recognizing old classes may lose due to severe overfitting. To tackle this problem, a few-shot class-incremental SAR ATR method, namely hierarchical embedding and incremental evolutionary network (HEIEN), is proposed in this paper. Firstly, a hierarchical embedding network and a hybrid distance-based classifier is constructed for basic feature extraction and classification. Then, in order to obtain more accurate decision boundaries, an adaptive class-incremental learning module is designed to adjust the weights of classifiers in all tasks by collecting context information from the past to the present. Finally, a pseudo incremental training strategy is designed to enable effective model training with only a few samples. Experimental results on the MSTAR benchmark data set have illustrated that HEIEN performs well with remarkable advantages in few-shot class-incremental SAR ATR tasks.

Incremental deep learning training approach for lesion detection and classification in mammograms

Recently, Deep Convolutional Neural Networks (DCNNs) have opened their ways into various medical image processing practices such as Computer-Aided Diagnosis (CAD) systems. Despite significant developments in CAD systems based on deep models, designing an efficient model, as well as a training strategy to cope with the shortage of medical images have yet to be addressed. To address current challenges, this paper presents a model including a hybrid DCNN, which takes advantage of various feature maps of different deep models and an incremental training algorithm. Also, a weighting Test Time Augmentation strategy is presented. Besides, the proposed work develops the Mask-RCNN to not only detect mass and calcification in mammography images, but also to classify normal images. Moreover, this work aims to benefit from a radiology specialist to compare with the performance of the proposed method. Illustrating the region of interest to explain how the model makes decisions is the other aim of the study to cover existing challenges among the stateof-the-art research works. The wide range of conducted quantitative and qualitative experiments suggest that the proposed method can classify breast X-ray images of the INbreast dataset to normal, mass, and calcification.

Multi-step ahead voltage prediction and voltage fault diagnosis based on gated recurrent unit neural network and incremental training

Accurate and early detection of voltage faults facilitates the driver and battery management system to take protective measures and reduce property damage and passenger injury. To identify the battery operation fault in a timely manner, this study develops an accurate multi-step voltage prediction and voltage fault diagnosis method based on gated recurrent unit neural network and incremental training. First, considering the impacts of drivers’ behaviors and vehicle states on battery performance under practical operations, a long-term operation dataset of electric scooters is acquired and established, and the Pearson correlation coefficient is applied to quantify these correlations. Then, the gated recurrent unit neural network, together with the multi-step ahead prediction scheme, is advanced to construct the voltage prediction model. Next, to effectively capture the performance variation of battery under complex dynamic operating environment, the incremental learning approach is developed to adaptively update the prediction model. Finally, the fault diagnosis strategy is proposed, with the combination of the voltage prediction model, to accurately detect battery faults of over-voltage, under-voltage, over-voltage change rate and poor consistency. The experimental validations highlight that the proposed method can predict the battery voltage 1 min in advance, and detect battery faults in real time with high accuracy.

Improved generalization of a plant-detection model for precision weed control

Lack of generalization in plant-detection models is one of the main challenges preventing the realization of autonomous weed-control systems. This paper investigates the effect of the train and test dataset distribution on the generalization error of a plant-detection model and uses incremental training to mitigate the said error. In this paper, we use the YOLOv3 object detector as plant-detection model. To train the model and test its generalization properties we used a broad dataset, consisting of 25 sub-datasets, sampled from multiple different geographic areas, soil types, cultivation conditions, containing variation in weeds, background vegetation, camera quality and variations in illumination. Using this dataset we evaluated the generalization error of a plant-detection model, assessed the effect of sampling training images from multiple arable fields on the generalization of our plant-detection model, we investigated the relation between the number of training images and the generalization of the plant-detection model and we applied incremental training to mitigate the generalization error of our plant-detection model on new arable fields. It was found that the average generalization error of our plant-detection model was 0.06 mAP. Increasing the number of sub-datasets for training, while keeping the total number of training images constant, increased the variation covered by the training set and improved the generalization of our plant-detection model. Adding more training images sampled from the same datasets increased the generalization further. However, this effect is limited and only holds when the new images cover new variation. Naively adding more images does not prepare the model for specific scenarios outside the training distribution. Using incremental training the model can be adapted to such scenarios and the generalization error can be mitigated. Depending on the discrepancy between the training set and the new field, finetuning on as little as 25 images can already mitigate the generalization error.

Dynamic Community Detection Method of a Social Network Based on Node Embedding Representation

The node embedding method enables network structure feature learning and representation for social network community detection. However, the traditional node embedding method only focuses on a node’s individual feature representation and ignores the global topological feature representation of the network. Traditional community detection methods cannot use the static node vector from the traditional node embedding method to calculate the dynamic features of the topological structure. In this study, an incremental dynamic community detection model based on a graph neural network node embedding representation is proposed, comprising the following aspects. A node embedding model based on influence random walk improves the information enrichment of the node feature vector representation, which improves the performance of the initial static community detection, whose results are used as the original structure of dynamic community detection. By combining a cohesion coefficient and ordinary modularity, a new modularity calculation method is proposed that uses an incremental training method to obtain node vector representation to detect a dynamic community from the perspectives of coarse- and fine-grained adjustments. A performance analysis based on two dynamic network datasets shows that the proposed method performs better than benchmark algorithms based on time complexity, community detection accuracy, and other indicators.

A Novel Approach for Semi-supervised Learning: Incremental Parallel Training with Cross-Validation (IPT-CV)

A Novel Approach for Semi-supervised Learning: Incremental Parallel Training with Cross-Validation (IPT-CV)

Virtual Reality in Vocational Training: A Study Demonstrating the Potential of a VR-based Vehicle Painting Simulator for Skills Acquisition in Apprenticeship Training

AbstractPrevious studies on Virtual Reality (VR)-enriched learning pointed out the advantages of immersive learning for the development of competencies. In the context of vocational education in vehicle painting, training opportunities are severely limited for many reasons. VR can be utilized to develop a comprehensive learning environment with authentic training tasks. Besides the need to train psychomotor skills, vehicle painting procedures are complex tasks requiring incremental training to develop knowledge, skills, and attitudes.This study aims to evaluate a VR training application for vehicle painting, focusing on the development of professional competencies regarding skills, knowledge, and attitudes.47 apprentices participated in the evaluation study. A VR-simulated painting booth was developed based on the 4 C/ID model by van Merriënboer, where they dealt with typical painting jobs on 3D workpieces (e.g., car wings, engine hood).Within the descriptive-inferential study, no significant differences between the types of competencies were revealed. The training application supports the acquisition of skills, knowledge, and attitudes equally. Further results regarding usability, cognitive load, etc., are promising.The essential finding of this study is that the VR training application is generally suitable for supporting craftsmanship within the field of vehicle painting. Since training opportunities for apprentices in this context are often rare, VR offers a unique solution especially for skills training if it follows a proven instructional model for the development of competencies.

Assessing the effects of data drift on the performance of machine learning models used in clinical sepsis prediction

BackgroundData drift can negatively impact the performance of machine learning algorithms (MLAs) that were trained on historical data. As such, MLAs should be continuously monitored and tuned to overcome the systematic changes that occur in the distribution of data. In this paper, we study the extent of data drift and provide insights about its characteristics for sepsis onset prediction. This study will help elucidate the nature of data drift for prediction of sepsis and similar diseases. This may aid with the development of more effective patient monitoring systems that can stratify risk for dynamic disease states in hospitals. MethodsWe devise a series of simulations that measure the effects of data drift in patients with sepsis, using electronic health records (EHR). We simulate multiple scenarios in which data drift may occur, namely the change in the distribution of the predictor variables (covariate shift), the change in the statistical relationship between the predictors and the target (concept shift), and the occurrence of a major healthcare event (major event) such as the COVID-19 pandemic. We measure the impact of data drift on model performances, identify the circumstances that necessitate model retraining, and compare the effects of different retraining methodologies and model architecture on the outcomes. We present the results for two different MLAs, eXtreme Gradient Boosting (XGB) and Recurrent Neural Network (RNN). ResultsOur results show that the properly retrained XGB models outperform the baseline models in all simulation scenarios, hence signifying the existence of data drift. In the major event scenario, the area under the receiver operating characteristic curve (AUROC) at the end of the simulation period is 0.811 for the baseline XGB model and 0.868 for the retrained XGB model. In the covariate shift scenario, the AUROC at the end of the simulation period for the baseline and retrained XGB models is 0.853 and 0.874 respectively. In the concept shift scenario and under the mixed labeling method, the retrained XGB models perform worse than the baseline model for most simulation steps. However, under the full relabeling method, the AUROC at the end of the simulation period for the baseline and retrained XGB models is 0.852 and 0.877 respectively. The results for the RNN models were mixed, suggesting that retraining based on a fixed network architecture may be inadequate for an RNN. We also present the results in the form of other performance metrics such as the ratio of observed to expected probabilities (calibration) and the normalized rate of positive predictive values (PPV) by prevalence, referred to as lift, at a sensitivity of 0.8. ConclusionOur simulations reveal that retraining periods of a couple of months or using several thousand patients are likely to be adequate to monitor machine learning models that predict sepsis. This indicates that a machine learning system for sepsis prediction will probably need less infrastructure for performance monitoring and retraining compared to other applications in which data drift is more frequent and continuous. Our results also show that in the event of a concept shift, a full overhaul of the sepsis prediction model may be necessary because it indicates a discrete change in the definition of sepsis labels, and mixing the labels for the sake of incremental training may not produce the desired results.

Robot path planner based on deep reinforcement learning and the seeker optimization algorithm

Path planning is one of the key technologies for mobile robot applications. However, the traditional robot path planner has a slow planning response, which leads to a long navigation completion time. In this paper, we propose a novel robot path planner (SOA+A2C) that produces global and local path planners with the seeker optimization algorithm (SOA) and the advantage actor-critic (A2C) algorithm, respectively. In addition, to solve the problems of poor convergence performance when training deep reinforcement learning (DRL) agents in complex path planning tasks and path redundancy when metaheuristic algorithms, such as SOA, are used for path planning, we propose the incremental map training method and path de-redundancy method. Simulation results show that first, the incremental map training method can improve the convergence performance of the DRL agent in complex path planning tasks. Second, the path de-redundancy method can effectively alleviate path redundancy without sacrificing the search capability of the metaheuristic algorithm. Third, the SOA+A2C path planner is superior to the Dijkstra & dynamic window approach (Dijkstra+DWA) and the Dijkstra & timed elastic band (Dijkstra+TEB) path planners provided by the robot operating system (ROS) in terms of path length, path planning response time, and navigation completion time. Therefore, the developed SOA+A2C path planner can serve as an effective tool for mobile robot path planning.

Ultra-Short-Term Load Dynamic Forecasting Method Considering Abnormal Data Reconstruction Based on Model Incremental Training

In order to reduce the influence of abnormal data on load forecasting effects and further improve the training efficiency of forecasting models when adding new samples to historical data set, an ultra-short-term load dynamic forecasting method considering abnormal data reconstruction based on model incremental training is proposed in this paper. Firstly, aiming at the abnormal data in ultra-short-term load forecasting, a load abnormal data processing method based on isolation forests and conditional adversarial generative network (IF-CGAN) is proposed. The isolation forest algorithm is used to accurately eliminate the abnormal data points, and a conditional generative adversarial network (CGAN) is constructed to interpolate the abnormal points. The load-influencing factors are taken as the condition constraints of the CGAN, and the weighted loss function is introduced to improve the reconstruction accuracy of abnormal data. Secondly, aiming at the problem of low model training efficiency caused by the new samples in the historical data set, a model incremental training method based on a bidirectional long short-term memory network (Bi-LSTM) is proposed. The historical data are used to train the Bi-LSTM, and the transfer learning is introduced to process the incremental data set to realize the adaptive and rapid adjustment of the model weight and improve the model training efficiency. Finally, the real power grid load data of a region in eastern China are used for simulation analysis. The calculation results show that the proposed method can reconstruct the abnormal data more accurately and improve the accuracy and efficiency of ultra-short-term load forecasting.

The potential role of selected body composition phenotypes in the interpretation of the exercise-induced components of the insulin-like growth factor systems

Introduction: Increased adiposity leads to impaired physiologic growth hormone secretion and low and high body mass index (BMI) values increase health risks. However, BMI only measures results in variations in fat-free mass (FFM), fat-free mass index (FFMI), normalized fat-free mass index (NFMI), and body fat mass (BFM). This study evaluated the insulin-like growth factor system responses to the given exercise and their interaction with the changes in BMI, FFM, FFMI, and NFFMI in healthy male participants. Material and methods: A randomized controlled trial with a parallel groups study design was used. Thirty healthy male participants (age: 21.33 ±1.24 years) were divided into three categories: high-intensity incremental (n = 12) and low-intensity constant (n = 12) cycling training groups and control group (n = 6). Training groups performed three times per week throughout eight weeks. VO2max, serum biomarkers, and neuromuscular performance were measured both during baseline and follow-up. Results: The changes in bioavailable IGF were not correlated with BMI (r = –.267), whereas they significantly positively correlated with BFM (r = .321), and inversely significantly correlated with FFM (r = –.472), FFMI (r = –.425), and NFFMI (r = –.379) after 8 weeks of exercise. For relative bioavailable IGF changes, FFM (r2 = 0.17), FFMI (r2 = 0.18), and NFFMI (r2 = 0.14) percent change explained nearly three times the variance as BMI percent change (r2 = 0.07). Conclusions: Increased bioavailable IGF-I suggests an increased anticatabolic effect and inverse interaction with body composition phenotypes following exercise. The partitioning of BMI into FFM, FFMI, and NFFMI rather than relying on sole measures of BMI seem to offer more precise results in the assessment of the interactions between the body composition, neuromuscular performance adjusted with body composition phenotypes, and training-induced changes in insulin-like growth factor system.

Incremental Learning of Deep Neural Network for Robust Vehicle Classification

Existing single-lane free flow (SLFF) tolling systems either heavily rely on contact-based treadle sensor to detect the number of vehicle wheels or manual operator to classify vehicles. While the former is susceptible to high maintenance cost due to wear and tear, the latter is prone to human error. This paper proposes a vision-based solution to SLFF vehicle classification by adapting a state-of-the-art object detection model as a backbone of the proposed framework and an incremental training scheme to train our VehicleDetNet in a continual manner to cater the challenging problem of continuous growing dataset in real-world environment. It involved four experiment set-ups where the first stage involved CUTe datasets. VehicleDetNet is utilized for the framework of vehicle detection, and it presents an anchorless network which enable the elimination of the bounding boxes of candidates’ anchors. The classification of vehicles is performed by detecting the vehicle’s location and inferring the vehicle’s class. We augment the model with a wheel detector and enumerator to add more robustness, showing improved performance. The proposed method was evaluated on live dataset collected from the Gombak toll plaza at Kuala Lumpur-Karak Expressway. The results show that within two months of observation, the mean accuracy increases from 87.3 % to 99.07 %, which shows the efficacy of our proposed method.

Knowledge distillation driven instance segmentation for grading prostate cancer

Prostate cancer (PCa) is one of the deadliest cancers in men, and identifying cancerous tissue patterns at an early stage can assist clinicians in timely treating the PCa spread. Many researchers have developed deep learning systems for mass-screening PCa. These systems, however, are commonly trained with well-annotated datasets in order to produce accurate results. Obtaining such data for training is often time and resource-demanding in clinical settings and can result in compromised screening performance. To address these limitations, we present a novel knowledge distillation-based instance segmentation scheme that allows conventional semantic segmentation models to perform instance-aware segmentation to extract stroma, benign, and the cancerous prostate tissues from the whole slide images (WSI) with incremental few-shot training. The extracted tissues are then used to compute majority and minority Gleason scores, which, afterward, are used in grading the PCa as per the clinical standards. The proposed scheme has been thoroughly tested on two datasets, containing around 10,516 and 11,000 WSI scans, respectively. Across both datasets, the proposed scheme outperforms state-of-the-art methods by 2.01% and 4.45%, respectively, in terms of the mean IoU score for identifying prostate tissues, and 10.73% and 11.42% in terms of F1 score for grading PCa according to the clinical standards. Furthermore, the applicability of the proposed scheme is tested under a blind experiment with a panel of expert pathologists, where it achieved a statistically significant Pearson correlation of 0.9192 and 0.8984 with the clinicians’ grading.

Flexible needle puncture path planning for liver tumors based on deep reinforcement learning

Objective. Minimally invasive surgery has been widely adopted in the treatment of patients with liver tumors. In liver tumor puncture surgery, an image-guided ablation needle for puncture surgery, which first reaches a target tumor along a predetermined path, and then ablates the tumor or injects drugs near the tumor, is often used to reduce patient trauma, improving the safety of surgery operations and avoiding possible damage to large blood vessels and key organs. In this paper, a path planning method for computer tomography (CT) guided ablation needle in liver tumor puncture surgery is proposed. Approach. Given a CT volume containing abdominal organs, we first classify voxels and optimize the number of voxels to reduce volume rendering pressure, then we reconstruct a multi-scale 3D model of the liver and hepatic vessels. Secondly, multiple entry points of the surgical path are selected based on the strong and weak constraints of clinical puncture surgery through multi-agent reinforcement learning. We select the optimal needle entry point based on the length measurement. Then, through the incremental training of the double deep Q-learning network (DDQN), the transmission of network parameters from the small-scale environment to the larger-scale environment is accomplished, and the optimal surgical path with more optimized details is obtained. Main results. To avoid falling into local optimum in network training, improve both the convergence speed and performance of the network, and maximize the cumulative reward, we train the path planning network on different scales 3D reconstructed organ models, and validate our method on tumor samples from public datasets. The scores of human surgeons verified the clinical relevance of the proposed method. Significance. Our method can robustly provide the optimal puncture path of flexible needle for liver tumors, which is expected to provide a reference for surgeons’ preoperative planning.