An algorithm for segmentation of lung ROI by mean-shift clustering combined with multi-scale HESSIAN matrix dot filtering

A new algorithm for segmentation of suspected lung ROI(regions of interest)by mean-shift clustering and multi-scale HESSIAN matrix dot filtering was proposed.Original image was firstly filtered by multi-scale HESSIAN matrix dot filters,round suspected nodular lesions in the image were enhanced,and linear shape regions of the trachea and vascular were suppressed.Then,three types of information,such as,shape filtering value of HESSIAN matrix,gray value,and spatial location,were introduced to feature space.The kernel function of mean-shift clustering was divided into product form of three kinds of kernel functions corresponding to the three feature information.Finally,bandwidths were calculated adaptively to determine the bandwidth of each suspected area,and they were used in mean-shift clustering segmentation.Experimental results show that by the introduction of HESSIAN matrix of dot filtering information to mean-shift clustering,nodular regions can be segmented from blood vessels,trachea,or cross regions connected to the nodule,non-nodular areas can be removed from ROIs properly,and ground glass object(GGO)nodular areas can also be segmented.For the experimental data set of 127 different forms of nodules,the average accuracy of the proposed algorithm is more than 90%.

GDMNet: A Unified Multi-Task Network for Panoptic Driving Perception

To enhance the efficiency and accuracy of environmental perception for autonomous vehicles,we propose GDMNet,a unified multi-task perception network for autonomous driving,capable of performing drivable area segmentation,lane detection,and traffic object detection.Firstly,in the encoding stage,features are extracted,and Generalized Efficient Layer Aggregation Network(GELAN)is utilized to enhance feature extraction and gradient flow.Secondly,in the decoding stage,specialized detection heads are designed;the drivable area segmentation head employs DySample to expand feature maps,the lane detection head merges early-stage features and processes the output through the Focal Modulation Network(FMN).Lastly,the Minimum Point Distance IoU(MPDIoU)loss function is employed to compute the matching degree between traffic object detection boxes and predicted boxes,facilitating model training adjustments.Experimental results on the BDD100K dataset demonstrate that the proposed network achieves a drivable area segmentation mean intersection over union(mIoU)of 92.2%,lane detection accuracy and intersection over union(IoU)of 75.3%and 26.4%,respectively,and traffic object detection recall and mAP of 89.7%and 78.2%,respectively.The detection performance surpasses that of other single-task or multi-task algorithm models.

Hybrid model enabling highly efficient follicular segmentation in thyroid cytopathological whole slide image

Background The prevalence of thyroid cancer is growing rapidly.Early and precise diagnosis is critical in thy-roid cancer caring.An automatic thyroid cancer diagnostic tool can be valuable to achieve early detection and diagnostic consistency.Only the follicular areas in the sample contain useful information to the thyroid cancer diagnosis based on fine needle aspiration(FNA).This study aimed to develop a highly efficient accurate method for follicular cell areas segmentation(FCAS)of thyroid cytopathological whole slide images(WSIs).Methods A total of 96 cell samples from July 2017 to July 2018 were collected in one hospital in Beijing,China.Forty-three WSIs were selected and manually labeled,including 17 cases of papillary thyroid carci-noma sample and 26 cases of benign sample.Six thousand and nine hundred cropped typical image patches(available on https://github.com/bupt-ai-cz/Hybrid-Model-Enabling-Highly-Efficient-Follicular-Segmentation)of 1024×1024 pixels from 13 large WSIs were used for patch-level model training and testing and all of the 13 large WSIs were papillary thyroid carcinoma samples.Thirty testing WSIs with an average size 36,217×29,400(from 10,240×10,240 to 81,920×61,440)were used to test the effectiveness of the hybrid model.Based on the traditional semantic segmentation model deeplabv3,we constructed a hybrid segmentation architecture by adding a classification branch into the segmentation scheme to improve efficiency.Accuracy was used to measure the performance of the classification model;pixel accuracy(pAcc),mean accuracy(mAcc),mean intersection over union(mIoU),and frequency weighted intersection over union(fwIoU)were used to measure the performance of the segmentation model,respectively.Results Using this method,up to 93%WSI segmentation time was reduced by skipping the colloidal areas and the blank background areas.The average processing time of 30 WSI was 49.49 s.On the patch dataset,this hybrid model might reach pAcc=98.65%,mAcc=85.60%,mIoU=79.61%,and fwIoU=97.54%.On the WSI dataset,this model might reach pAcc=99.30%,mAcc=68.94%,mIoU=58.21%,and fwIoU=99.50%.Conclusion The proposed hybrid method might significantly improve previous solutions and achieve the superior performance of efficiency and accuracy.

YOLOP:You Only Look Once for Panoptic Driving Perception

A panoptic driving perception system is an essential part of autonomous driving.A high-precision and real-time perception system can assist the vehicle in making reasonable decisions while driving.We present a panoptic driving perception network(you only look once for panoptic(YOLOP))to perform traffic object detection,drivable area segmentation,and lane detection simultaneously.It is composed of one encoder for feature extraction and three decoders to handle the specific tasks.Our model performs extremely well on the challenging BDD100K dataset,achieving state-of-the-art on all three tasks in terms of accuracy and speed.Besides,we verify the effectiveness of our multi-task learning model for joint training via ablative studies.To our best knowledge,this is the first work that can process these three visual perception tasks simultaneously in real-time on an embedded device Jetson TX2(23 FPS),and maintain excellent accuracy.To facilitate further research,the source codes and pre-trained models are released at https://github.com/hustvl/YOLOP.