Medical Image

Select

Multi-Model Ensemble Deep Learning Method to Diagnose COVID-19 Using Chest Computed Tomography Images

WANG Zhiming(王志明), DONG Jingjing (董静静), ZHANG Junpeng∗ (张军鹏)

J Shanghai Jiaotong Univ Sci 2022, 27 (1): 70-80. DOI: 10.1007/s12204-021-2392-3

Abstract （679）

PDF（pc）（1418KB）（194）

Save

Deep learning based analyses of computed tomography (CT) images contribute to automated diagnosis of COVID-19, and ensemble learning may commonly provide a better solution. Here, we proposed an ensemble learning method that integrates several component neural networks to jointly diagnose COVID-19. Two ensemble strategies are considered: the output scores of all component models that are combined with the weights adjusted adaptively by cost function back propagation; voting strategy. A database containing 8 347 CT slices of COVID- 19, common pneumonia and normal subjects was used as training and testing sets. Results show that the novel method can reach a high accuracy of 99.37% (recall: 0.998 1; precision: 0.989 3), with an increase of about 7% in comparison to single-component models. And the average test accuracy is 95.62% (recall: 0.958 7; precision: 0.955 9), with a corresponding increase of 5.2%. Compared with several latest deep learning models on the identical test set, our method made an accuracy improvement up to 10.88%. The proposed method may be a promising solution for the diagnosis of COVID-19.

Reference | Related Articles | Metrics | Comments（0）

Select

COVID-19 Interpretable Diagnosis Algorithm Based on a Small Number of Chest X-Ray Samples

BU Ran (卜冉), XIANG Wei∗ (向伟), CAO Shitong (曹世同)

J Shanghai Jiaotong Univ Sci 2022, 27 (1): 81-89. DOI: 10.1007/s12204-021-2393-2

Abstract （777）

PDF（pc）（1470KB）（288）

Save

The COVID-19 medical diagnosis method based on individual’s chest X-ray (CXR) is achieved difficultly in the initial research, owing to difficulties in identifying CXR data of COVID-19 individuals. At the beginning of the study, infected individuals’ CXRs were scarce. The combination of artificial intelligence (AI) and medical diagnosis has been advanced and popular. To solve the difficulties, the interpretability analysis of AI model was used to explore the pathological characteristics of CXR samples infected with COVID-19 and assist in medical diagnosis. The dataset was expanded by data augmentation to avoid overfitting. Transfer learning was used to test different pre-trained models and the unique output layers were designed to complete the model training with few samples. In this study, the output results of four pre-trained models in three different output layers were compared, and the results after data augmentation were compared with the results of the original dataset. The control variable method was used to conduct independent tests of 24 groups. Finally, 99.23% accuracy and 98% recall rate were obtained, and the visual results of CXR interpretability analysis were displayed. The network of COVID-19 interpretable diagnosis algorithm has the characteristics of high generalization and lightweight. It can be quickly applied to other urgent tasks with insufficient experimental data. At the same time, interpretability analysis brings new possibilities for medical diagnosis.

Reference | Related Articles | Metrics | Comments（0）

Select

Application of Deep Learning Method on Ischemic Stroke Lesion Segmentation

ZHANG Yue (张月), LIU Shijie (刘世界), LI Chunlai (李春来), WANG Jianyu (王建宇)

J Shanghai Jiaotong Univ Sci 2022, 27 (1): 99-111. DOI: 10.1007/s12204-021-2273-9

Abstract （830）

PDF（pc）（944KB）（222）

Save

Although deep learning methods have been widely applied in medical image lesion segmentation, it is still challenging to apply them for segmenting ischemic stroke lesions, which are different from brain tumors in lesion characteristics, segmentation difficulty, algorithm maturity, and segmentation accuracy. Three main stages are used to describe the manifestations of stroke. For acute ischemic stroke, the size of the lesions is similar to that of brain tumors, and the current deep learning methods have been able to achieve a high segmentation accuracy. For sub-acute and chronic ischemic stroke, the segmentation results of mainstream deep learning algorithms are still unsatisfactory as lesions in these stages are small and diffuse. By using three scientific search engines including CNKI, Web of Science and Google Scholar, this paper aims to comprehensively understand the state-of-the-art deep learning algorithms applied to segmenting ischemic stroke lesions. For the first time, this paper discusses the current situation, challenges, and development directions of deep learning algorithms applied to ischemic stroke lesion segmentation in different stages. In the future, a system that can directly identify different stroke stages and automatically select the suitable network architecture for the stroke lesion segmentation needs to be proposed.

Reference | Related Articles | Metrics | Comments（0）

Select

SeRN: A Two-Stage Framework of Registration for Semi-Supervised Learning for Medical Images

JIA Dengqiang* (贾灯强), LUO Xinzhe (罗鑫喆), DING Wangbin (丁王斌),HUANG Liqin (黄立勤), ZHUANG Xiahai (庄吓海)

J Shanghai Jiaotong Univ Sci 2022, 27 (2): 176-189. DOI: 10.1007/s12204-021-2383-4

Abstract （833）

PDF（pc）（2406KB）（269）

Save

Significant breakthroughs in medical image registration have been achieved using deep neural networks (DNNs). However, DNN-based end-to-end registration methods often require large quantities of data or adequate annotations for training. To leverage the intensity information of abundant unlabeled images, unsupervised registration methods commonly employ intensity-based similarity measures to optimize the network parameters.However, finding a sufficiently robust measure can be challenging for specific registration applications. Weakly supervised registration methods use anatomical labels to estimate the deformation between images. High-level structural information in label images is more reliable and practical for estimating the voxel correspondence of anatomic regions of interest between images, whereas label images are extremely difficult to collect. In this paper, we propose a two-stage semi-supervised learning framework for medical image registration, which consists of unsupervised and weakly supervised registration networks. The proposed semi-supervised learning framework is trained with intensity information from available images, label information from a relatively small number of labeled images and pseudo-label information from unlabeled images. Experimental results on two datasets (cardiac and abdominal images) demonstrate the efficacy and efficiency of this method in intra- and inter-modality medical image registrations, as well as its superior performance when a vast amount of unlabeled data and a small set of annotations are available. Our code is publicly available at https://github.com/jdq818/SeRN.

Reference | Related Articles | Metrics | Comments（0）

Select

Interactive Liver Segmentation Algorithm Based on Geodesic Distance and V-Net

KANG Jie* (亢洁), DING Jumin (丁菊敏), LEI Tao (雷涛),FENG Shujie (冯树杰), LIU Gang (刘港)

J Shanghai Jiaotong Univ Sci 2022, 27 (2): 190-201. DOI: 10.1007/s12204-021-2379-0

Abstract （683）

PDF（pc）（3921KB）（228）

Save

Convolutional neural networks (CNNs) are prone to mis-segmenting image data of the liver when the background is complicated, which results in low segmentation accuracy and unsuitable results for clinical use. To address this shortcoming, an interactive liver segmentation algorithm based on geodesic distance and V-net is proposed. The three-dimensional segmentation network V-net adequately considers the characteristics of the spatial context information to segment liver medical images and obtain preliminary segmentation results. An artificial algorithm based on geodesic distance is used to form artificial hard constraints to modify the image,and the superpixel piece created by the watershed algorithm is introduced as a sample point for operation, which significantly improves the efficiency of segmentation. Results from simulation of the liver tumor segmentation challenge (LiTS) dataset show that this algorithm can effectively refine the results of automatic liver segmentation,reduce user intervention, and enable a fast, interactive liver image segmentation that is convenient for doctors.

Reference | Related Articles | Metrics | Comments（0）

Select

Survey of EIT Image Reconstruction Algorithms

ZHANG Mingzhu(张明珠), MA Yixin* (马艺馨), HUANG Ningning (黄宁宁), GE Hao (葛浩)

J Shanghai Jiaotong Univ Sci 2022, 27 (2): 211-218. DOI: 10.1007/s12204-021-2333-1

Abstract （610）

PDF（pc）（272KB）（185）

Save

With the recent promotion of clinical applications of electrical impedance tomography (EIT) technology,more scholars have begun studying EIT technology. Although the principle of EIT technology seems simple,EIT image reconstruction is a non-linear and ill-posed problem that is quite difficult to solve because of its soft field characteristics and the inhomogeneous distribution of its sensitive field. What’s more, the EIT reconstruction algorithm requires further improvements in robustness, clarity, etc. The image-reconstruction algorithm and image quality are among the key challenges in the application of EIT technology; thus, more research is urgently needed to improve the performance of EIT technology and use it to solve a larger variety of clinical problems. In this paper, we pay special attention to the latest advances in the study of EIT image-reconstruction algorithms to provide a convenient reference for EIT beginners and researchers who are newly involved in research on EIT image reconstruction.

Reference | Related Articles | Metrics | Comments（0）

Select

Evaluation Value of Intravoxel Incoherent Motion Diffusion-Weighted Imaging on Early Efficacy of Magnetic Resonance-Guided High-Intensity Focused Ultrasound Ablation for Uterine Adenomyoma

TANG Na (唐纳), GU Jianjun (顾坚骏), YIN Xiaorui (尹肖睿), YU Rongjiang (虞容江),XU Yuantao (徐元涛), LI Xiang (李想), WANG Han* (王悍)

J Shanghai Jiaotong Univ Sci 2022, 27 (2): 226-230. DOI: 10.1007/s12204-022-2405-x

Abstract （775）

PDF（pc）（549KB）（196）

Save

To investigate the evaluation value of intravoxel incoherent motion diffusion-weighted imaging (IVIMDWI) on the early efficacy of magnetic resonance-guided high-intensity focused ultrasound (MRgFUS) ablation for uterine adenomyoma. The clinical and magnetic resonance imaging (MRI) data of 36 patients with uterine adenomyoma before and after MRgFUS treatment in our hospital from January 2018 to December 2018 were retrospectively analyzed. All the 36 patients underwent MRI examination one day before operation and immediately after operation using GE Discovery MR750 3.0T MRI, including conventional sequences (T1WI, T2WI,and T2 fat suppression sequences) plain scan, IVIM-DWI sequences with 9 b values, and contrast enhanced-MRI sequences. The IVIM-DWI quantitative parameters (true diffusion coefficient D, perfusion related diffusion coefficient D?, and perfusion fraction f) of double-exponential model were obtained by using GE ADW 4.7 functool,a postprocessor. SPSS 24.0 software was used to analyze the difference in parameter between the ablation and non-ablation areas of uterine adenomyoma. DWI signal in the ablation area of uterine adenomyoma was increased,and manifested as heterogeneous diffuse high signal, with low central signal and high edge signal. Values of D, D? and f in the ablation area of uterine adenomyoma were significantly lower than those in the non-ablation area,and there was statistical difference between the two (P <0.05). The areas under receiver operating characteristic (ROC) curve of D, D? and f values in the ablation area of uterine adenomyoma were 0.854, 0.898 and 0.924,respectively; the optimal thresholds for the diagnosis of ablation area of uterine adenomyoma were 0.81 × 10 ?3 mm2/s, 4.99×10 ?3 mm2/s and 0.24, respectively; the diagnostic sensitivity was 80.6%, 72.2% and 94.4%, respectively; and the specificity was 91.7%, 97.2% and 94.4%, respectively. IVIM-DWI has a certain clinical value in the evaluation on early efficacy of MRgFUS ablation of uterine adenomyosis.

Reference | Related Articles | Metrics | Comments（0）

Select

Deformable Registration Algorithm via Non-subsampled Contourlet Transform and Saliency Map

CHANG Qing∗ (常青), YANG Wenyou (杨文友), CHEN Lanlan (陈兰岚)

J Shanghai Jiaotong Univ Sci 2022, 27 (4): 452-462. DOI: 10.1007/s12204-022-2428-3

Abstract （495）

PDF（pc）（4867KB）（187）

Save

Medical image registration is widely used in image-guided therapy and image-guided surgery to esti- mate spatial correspondence between planning and treatment images. However, most methods based on intensity have the problems of matching ambiguity and ignoring the influence of weak correspondence areas on the overall registration. In this study, we propose a novel general-purpose registration algorithm based on free-form defor- mation by non-subsampled contourlet transform and saliency map, which can reduce the matching ambiguities and maintain the topological structure of weak correspondence areas. An optimization method based on Markov random fields is used to optimize the registration process. Experiments on four public datasets from brain, car- diac, and lung have demonstrated the general applicability and the accuracy of our algorithm compared with two state-of-the-art methods.

Reference | Related Articles | Metrics | Comments（0）

Select

Breast Pathological Image Classification Based on VGG16 Feature Concatenation

LIU Min (刘敏), YI Ming (易鸣), WU Minghu∗ (武明虎), WANG Juan (王娟), HE Yu (何宇)

J Shanghai Jiaotong Univ Sci 2022, 27 (4): 473-484. DOI: 10.1007/s12204-021-2398-x

Abstract （920）

PDF（pc）（5914KB）（239）

Save

Breast cancer is one of the malignancies that endanger women’s health all over the world. Considering that there is some noise and edge blurring in breast pathological images, it is easier to extract shallow features of noise and redundant information when VGG16 network is used, which is affected by its relative shallow depth and small convolution kernel. To improve the pathological diagnosis of breast cancers, we propose a classification method for benign and malignant tumors in the breast pathological images which is based on feature concatenation of VGG16 network. First, in order to improve the problems of small dataset size and unbalanced data samples, the original BreakHis dataset is processed by data augmentation technologies, such as geometric transformation and color enhancement. Then, to reduce noise and edge blurring in breast pathological images, we perform bilateral filtering and denoising on the original dataset and sharpen the edge features by Sobel operator, which makes the extraction of shallow features by VGG16 model more accurate. Based on transfer learning, the network model trained with the expanded dataset is called VGG16-1, and another model trained with the image denoising and sharpening and mixed with the original dataset is called VGG16-2. The features extracted by VGG16-1 and VGG16-2 are concatenated, and then classified by support vector machine. The final experimental results show that the average accuracy is 98.44%, 98.89%, 98.30% and 97.47%, respectively, when the proposed method is tested with the breast pathological images of 40×, 100×, 200× and 400× on BreakHis dataset.

Reference | Related Articles | Metrics | Comments（0）

Select

USSL Net: Focusing on Structural Similarity with Light U-Structure for Stroke Lesion Segmentation

JIANG Zhiguo (蒋志国), CHANG Qing∗ (常青)

J Shanghai Jiaotong Univ Sci 2022, 27 (4): 485-497. DOI: 10.1007/s12204-022-2412-y

Abstract （766）

PDF（pc）（1121KB）（195）

Save

Automatic segmentation of ischemic stroke lesions from computed tomography (CT) images is of great significance for identifying and curing this life-threatening condition. However, in addition to the problem of low image contrast, it is also challenged by the complex changes in the appearance of the stroke area and the difficulty in obtaining image data. Considering that it is difficult to obtain stroke data and labels, a data enhancement algorithm for one-shot medical image segmentation based on data augmentation using learned transformation was proposed to increase the number of data sets for more accurate segmentation. A deep convolutional neural network based algorithm for stroke lesion segmentation, called structural similarity with light U-structure (USSL) Net, was proposed. We embedded a convolution module that combines switchable normalization, multi-scale convolution and dilated convolution in the network for better segmentation performance. Besides, considering the strong structural similarity between multi-modal stroke CT images, the USSL Net uses the correlation maximized structural similarity loss (SSL) function as the loss function to learn the varying shapes of the lesions. The experimental results show that our framework has achieved results in the following aspects. First, the data obtained by adding our data enhancement algorithm is better than the data directly segmented from the multi- modal image. Second, the performance of our network model is better than that of other models for stroke segmentation tasks. Third, the way SSL functioned as a loss function is more helpful to the improvement of segmentation accuracy than the cross-entropy loss function.

Reference | Related Articles | Metrics | Comments（0）

Select

Improving Colonoscopy Polyp Detection Rate Using Semi-Supervised Learning

YAO Leyul (姚乐宇)，HE Fan1,3 (何凡), PENG Haixia2* (彭海霞), WANG Xiaofeng2 (王晓峰)，ZHOU Lu2(周璐), HUANG Xiaolin1,3* (黄晓霖)

J Shanghai Jiaotong Univ Sci 2023, 28 (4): 441-. DOI: 10.1007/s12204-022-2519-1

Abstract （703）

PDF（pc）（497KB）（373）

Save

Colorectal cancer is one of the biggest health threats to humans and takes thousands of lives every year.Colonoscopy is the gold standard in clinical practice to inspect the intestinal wall, detect polyps and remove polypsin early stages, preventing polyps from becoming malignant and forming colorectal cancer instances. In recentyears, computer-aided polyp detection systems have been widely used in colonoscopies to improve the qualityof colonoscopy examination and increase the polyp detection rate. Currently, the most efficient computer-aidedsystems are built with machine learning methods. However, developing such a computer-aided detection systemrequires experienced doctors to label a large number of image data from colonoscopy videos, which is extremelytime-consuming, laborious and expensive. One possible solution is to adopt a semi-supervised learning, which canbuild a detection system on a dataset where part of its data is not necessary to be labeled. In this paper, on thebasis of state-of-the-art object detection method and semi-supervised learning technique, we design and implementa semi-supervised colonoscopy polyp detection system containing four main steps: running standard supervisedtraining with all labeled data; running inference on unlabeled data to obtain pseudo labels; applying a set ofstrong augmentation to both unlabeled data and pseudo label; combining labeled data, and unlabeled data withits pseudo labels to retrain the detector. The semi-supervised learning system is evaluated both on public datasetand our original private dataset and proves its effectiveness. Also, the inference speed of the semi-supervisedlearning system can meet the requirement of real-time operation.

Reference | Related Articles | Metrics | Comments（0）

Select

Ensemble of Two-Path Capsule Networks for Diagnosis of Turner Syndrome Using Global-Local Facial Images

LIU Lu (刘璐)

J Shanghai Jiaotong Univ Sci 2023, 28 (4): 459-. DOI: 10.1007/s12204-022-2491-9

Abstract （720）

PDF（pc）（545KB）（225）

Save

Turner syndrome (TS) is a chromosomal disorder disease that only affects the growth of female patients. Prompt diagnosis is of high significance for the patients. However, clinical screening methods are time-consuming and cost-expensive. Some researchers used machine learning-based methods to detect TS, the performance of which needed to be improved. Therefore, we propose an ensemble method of two-path capsule networks (CapsNets) for detecting TS based on global-local facial images. Specifically, the TS facial images are preprocessed and segmented into eight local parts under the direction of physicians; then, nine two-path CapsNets are respectively trained using the complete TS facial images and eight local images, in which the few-shot learning is utilized to solve the problem of limited data; finally, a probability-based ensemble method is exploited to combine nine classifiers for the classification of TS. By studying base classifiers, we find two meaningful facial areas are more related to TS patients, i.e., the parts of eyes and nose. The results demonstrate that the proposed model is effective for the TS classification task, which achieves the highest accuracy of 0.924 1.

Reference | Related Articles | Metrics | Comments（0）

Select

Dlung: Unsupervised Few-Shot Diffeomorphic Respiratory Motion Modeling

CHEN Peizhi1,2* (陈培芝), GUO Yifan1 (郭逸凡)，WANG Dahan1,2 (王大寒), CHEN Chinling1,3,4* (陈金铃)

J Shanghai Jiaotong Univ Sci 2023, 28 (4): 536-. DOI: 10.1007/s12204-022-2525-3

Abstract （509）

PDF（pc）（1720KB）（344）

Save

Lung image registration plays an important role in lung analysis applications, such as respiratory motion modeling. Unsupervised learning-based image registration methods that can compute the deformation without the requirement of supervision attract much attention. However, it is noteworthy that they have two drawbacks: they do not handle the problem of limited data and do not guarantee diffeomorphic (topologypreserving) properties, especially when large deformation exists in lung scans. In this paper, we present an unsupervised few-shot learning-based diffeomorphic lung image registration, namely Dlung. We employ fine-tuning techniques to solve the problem of limited data and apply the scaling and squaring method to accomplish the diffeomorphic registration. Furthermore, atlas-based registration on spatio-temporal (4D) images is performed and thoroughly compared with baseline methods. Dlung achieves the highest accuracy with diffeomorphic properties. It constructs accurate and fast respiratory motion models with limited data. This research extends our knowledge of respiratory motion modeling.

Reference | Related Articles | Metrics | Comments（0）

Select

Time-Resolved Imaging in Short-Wave Infrared Region

XU Yang (徐杨), LI Wanwan∗ (李万万)

J Shanghai Jiaotong Univ Sci 2024, 29 (1): 29-36. DOI: 10.1007/s12204-022-2547-x

Abstract （1411）

PDF（pc）（810KB）（381）

Save

Compared with the conventional first near-infrared (NIR-I, 700—900 nm) window, the short-wave infrared region (SWIR, 900—1 700 nm) possesses the merits of the increasing tissue penetration depths and the suppression of scattering background, leading to great potential for in vivo imaging. Based on the limitations of the common spectral domain, and the superiority of the time-dimension, time-resolved imaging eliminates the auto-fluorescence in the biological tissue, thus supporting higher signal-to-noise ratio and sensitivities. The imaging technique is not affected by the difference in tissue composition or thickness and has the practical value of quantitative in vivo detection. Almost all the relevant time-resolved imaging was carried out around lanthanide-doped upconversion nanomaterials, owing to the advantages of ultralong luminescence lifetime, excellent photostability, controllable morphology, easy surface modification and various strategies of regulating lifetime. Therefore, this review presents the research progress of SWIR time-resolved imaging technology based on nanomaterials doped with lanthanide ions as luminescence centers in recent years.

Reference | Related Articles | Metrics | Comments（0）

Select

Retinal Vessel Segmentation via Adversarial Learning and Iterative Refinement

GU Wen (顾闻), XU Yi∗ (徐奕)

J Shanghai Jiaotong Univ Sci 2024, 29 (1): 73-80. DOI: 10.1007/s12204-022-2479-5

Abstract （989）

PDF（pc）（914KB）（225）

Save

Retinal vessel segmentation is a challenging medical task owing to small size of dataset, micro blood vessels and low image contrast. To address these issues, we introduce a novel convolutional neural network in this paper, which takes the advantage of both adversarial learning and recurrent neural network. An iterative design of network with recurrent unit is performed to refine the segmentation results from input retinal image gradually. Recurrent unit preserves high-level semantic information for feature reuse, so as to output a sufficiently refined segmentation map instead of a coarse mask. Moreover, an adversarial loss is imposing the integrity and connectivity constraints on the segmented vessel regions, thus greatly reducing topology errors of segmentation. The experimental results on the DRIVE dataset show that our method achieves area under curve and sensitivity of 98.17% and 80.64%, respectively. Our method achieves superior performance in retinal vessel segmentation compared with other existing state-of-the-art methods.

Reference | Related Articles | Metrics | Comments（0）

Select

Unsupervised Oral Endoscope Image Stitching Algorithm

HUANG Rong (黄荣), CHANG Qing^∗ (常青), ZHANG Yang （张扬）

J Shanghai Jiaotong Univ Sci 2024, 29 (1): 81-90. DOI: 10.1007/s12204-022-2513-7

Abstract （908）

PDF（pc）（5774KB）（294）

Save

Oral endoscope image stitching algorithm is studied to obtain wide-field oral images through registration and stitching, which is of great significance for auxiliary diagnosis. Compared with natural images, oral images have lower textures and fewer features. However, traditional feature-based image stitching methods rely heavily on feature extraction quality, often showing an unsatisfactory performance when stitching images with few features. Moreover, due to the hand-held shooting, there are large depth and perspective disparities between the captured images, which also pose a challenge to image stitching. To overcome the above problems, we propose an unsupervised oral endoscope image stitching algorithm based on the extraction of overlapping regions and the loss of deep features. In the registration stage, we extract the overlapping region of the input images by sketching polygon intersection for feature points screening and estimate homography from coarse to fine on a three-layer feature pyramid structure. Moreover, we calculate loss using deep features instead of pixel values to emphasize the importance of depth disparities in homography estimation. Finally, we reconstruct the stitched images from feature to pixel, which can eliminate artifacts caused by large parallax. Our method is compared with both feature-based and previous deep-based methods on the UDIS-D dataset and our oral endoscopy image dataset. The experimental results show that our algorithm can achieve higher homography estimation accuracy, and better visual quality, and can be effectively applied to oral endoscope image stitching.

Reference | Related Articles | Metrics | Comments（0）

Select

Medical Image Encryption Based on Josephus Traversing and Hyperchaotic Lorenz System

YANG Na (杨娜), ZHANG Shuxia (张淑霞), BAI Mudan (白牡丹), LI Shanshan^∗ (李珊珊)

J Shanghai Jiaotong Univ Sci 2024, 29 (1): 91-108. DOI: 10.1007/s12204-022-2555-x

Abstract （852）

PDF（pc）（8082KB）（556）

Save

This study proposes a new medical image encryption scheme based on Josephus traversing and hyperchaotic Lorenz system. First, a chaotic sequence is generated through hyperchaotic system. This hyperchaotic sequence is used in the scrambling and diffusion stages of the algorithm. Second, in the scrambling process, the image is initially confused by Josephus scrambling, and then the image is further confused by Arnold map. Finally, generated hyperchaos sequence and exclusive OR operation is used for the image to carry on the positive and reverse diffusion to change the pixel value of the image and further hide the effective information of the image. In addition, the information of the plaintext image is used to generate keys used in the algorithm, which increases the ability of resisting plaintext attack. Experimental results and security analysis show that the scheme can effectively hide plaintext image information according to the characteristics of medical images, and is resistant to common types of attacks. In addition, this scheme performs well in the experiments of robustness, which shows that the scheme can solve the problem of image damage in telemedicine. It has a positive significance for the future research.

Reference | Related Articles | Metrics | Comments（0）

Select

Ensemble Attention Guided Multi-SEANet Trained with Curriculum Learning for Noninvasive Prediction of Gleason Grade Groups from MRI

SHEN Ao^1,2‡ (沈傲), HU Jisu^2,3‡ (胡冀苏), JIN Pengfei⁴ (金鹏飞), ZHOU Zhiyong² (周志勇), QIAN Xusheng^2,3 (钱旭升), ZHENG Yi² (郑毅), BAO Jie⁴ (包婕), WANG Ximing^4∗ (王希明), DAI Yakang^1,2∗ (戴亚康)

J Shanghai Jiaotong Univ Sci 2024, 29 (1): 109-119. DOI: 10.1007/s12204-022-2502-x

Abstract （1099）

PDF（pc）（1407KB）（267）

Save

The Gleason grade group (GG) is an important basis for assessing the malignancy of prostate cancer, but it requires invasive biopsy to obtain pathology. To noninvasively evaluate GG, an automatic prediction method is proposed based on multi-scale convolutional neural network of the ensemble attention module trained with curriculum learning. First, a lesion-attention map based on the image of the region of interest is proposed in combination with the bottleneck attention module to make the network more focus on the lesion area. Second, the feature pyramid network is combined to make the network better learn the multi-scale information of the lesion area. Finally, in the network training, a curriculum based on the consistency gap between the visual evaluation and the pathological grade is proposed, which further improves the prediction performance of the network. Experimental results show that the proposed method is better than the traditional network model in predicting GG performance. The quadratic weighted Kappa is 0.471 1 and the positive predictive value for predicting clinically significant cancer is 0.936 9.

Reference | Related Articles | Metrics | Comments（0）

Select

Weighted Heterogeneous Graph-Based Incremental Automatic Disease Diagnosis Method

TIAN Yuanyuan (田圆圆), JIN Yanrui (金衍瑞), LI Zhiyuan (李志远), LIU Jinlei (刘金磊), LIU Chengliang^∗ (刘成良)

J Shanghai Jiaotong Univ Sci 2024, 29 (1): 120-130. DOI: 10.1007/s12204-022-2537-z

Abstract （894）

PDF（pc）（1081KB）（230）

Save

The objective of this study is to construct a multi-department symptom-based automatic diagnosis model. However, it is difficult to establish a model to classify plenty of diseases and collect thousands of diseasesymptom datasets simultaneously. Inspired by the thought of “knowledge graph is model”, this study proposes to build an experience-infused knowledge model by continuously learning the experiential knowledge from data, and incrementally injecting it into the knowledge graph. Therefore, incremental learning and injection are used to solve the data collection problem, and the knowledge graph is modeled and containerized to solve the large-scale multi-classification problems. First, an entity linking method is designed and a heterogeneous knowledge graph is constructed by graph fusion. Then, an adaptive neural network model is constructed for each dataset, and the data is used for statistical initialization and model training. Finally, the weights and biases of the learned neural network model are updated to the knowledge graph. It is worth noting that for the incremental process, we consider both the data and class increments. We evaluate the diagnostic effectiveness of the model on the current dataset and the anti-forgetting ability on the historical dataset after class increment on three public datasets. Compared with the classical model, the proposed model improves the diagnostic accuracy of the three datasets by 5%, 2%, and 15% on average, respectively. Meanwhile, the model under incremental learning has a better ability to resist forgetting.

Reference | Related Articles | Metrics | Comments（0）

Select

TshFNA-Examiner: A Nuclei Segmentation and Cancer Assessment Framework for Thyroid Cytology Image

KE Jing¹(柯晶), ZHU Junchao² (朱俊超), YANG Xin¹(杨鑫), ZHANG Haolin³ (张浩林), SUN Yuxiang¹(孙宇翔), WANG Jiayi¹(王嘉怡), LU Yizhou⁴(鲁亦舟), SHEN Yiqing⁵(沈逸卿), LIU Sheng⁶(刘晟), JIANG Fusong⁷(蒋伏松), HUANG Qin⁸(黄琴)

J Shanghai Jiaotong Univ Sci 2024, 29 (6): 945-957. DOI: 10.1007/s12204-024-2743-y

Abstract （807）

PDF（pc）（2836KB）（461）

Save

Examining thyroid fine-needle aspiration (FNA) can grade cancer risks, derive prognostic information, and guide follow-up care or surgery. The digitization of biopsy and deep learning techniques has recently enabled computational pathology. However, there is still lack of systematic diagnostic system for the complicated gigapixel cytopathology images, which can match physician-level basic perception. In this study, we design a deep learning framework, thyroid segmentation and hierarchy fine-needle aspiration (TshFNA)-Examiner to quantitatively profile the cancer risk of a thyroid FNA image. In the TshFNA-Examiner, cellular-intensive areas strongly correlated with diagnostic medical information are detected by a nuclei segmentation neural network; cell-level image patches are catalogued following The Bethesda System for Reporting Thyroid Cytopathology (TBSRTC) system, by a classification neural network which is further enhanced by leveraging unlabeled data. A cohort of 333 thyroid FNA cases collected from 2019 to 2022 from I to VI is studied, with pixel-wise and image-wise image patches annotated. Empirically, TshFNA-Examiner is evaluated with comprehensive metrics and multiple tasks to demonstrate its superiority to state-of-the-art deep learning approaches. The average performance of cellular area segmentation achieves a Dice of 0.931 and Jaccard index of 0.871. The cancer risk classifier achieves a macro-F1-score of 0.959, macro-AUC of 0.998, and accuracy of 0.959 following TBSRTC. The corresponding metrics can be enhanced to a macro-F1-score of 0.970, macro-AUC of 0.999, and accuracy of 0.970 by leveraging informative unlabeled data. In clinical practice, TshFNA-Examiner can help cytologists to visualize the output of deep learning networks in a convenient way to facilitate making the final decision.

Reference | Related Articles | Metrics | Comments（0）

Select

Video-Based Detection of Epileptic Spasms in IESS: Modeling, Detection, and Evaluation

DING Lihui^{1, 2}(丁黎辉), FU Lijun^{1, 3} (付立军), YANG Guang⁴(杨光), WAN Lin^{4, 5} (万林), CHANG Zhijun⁷(常志军)

J Shanghai Jiaotong Univ Sci 2025, 30 (1): 1-9. DOI: 10.1007/s12204-024-2789-x

Abstract （1349）

PDF（pc）（712KB）（222）

Save

Behavioral scoring based on clinical observations remains the gold standard for screening, diagnosing,and evaluating infantile epileptic spasm syndrome (IESS). The accurate identification of seizures is crucial for clinical diagnosis and assessment. In this study, we propose an innovative seizure detection method based on video feature recognition of patient spasms. To capture the temporal characteristics of the spasm behavior presented in the videos effectively, we incorporate asymmetric convolution and convolution–batch normalization–ReLU (CBR) modules. Specifically within the 3D-ResNet residual blocks, we split the larger convolutional kernels into two asymmetric 3D convolutional kernels. These kernels are connected in series to enhance the ability of the convolutional layers to extract key local features, both horizontally and vertically. In addition, we introduce a 3D convolutional block attention module to enhance the spatial correlations between video frame channels efficiently. To improve the generalization ability, we design a composite loss function that combines cross-entropy loss with triplet loss to balance the classification and similarity requirements. We train and evaluate our method using the PLA IESS-VIDEO dataset, achieving an average seizure recognition accuracy of 90.59%, precision of 90.94%, and recall of 87.64%. To validate its generalization capability further, we conducted external validation using six different patient monitoring videos compared with assessments by six human experts from various medical centers. The final test results demonstrate that our method achieved a recall of 0.647 6, surpassing the average level achieved by human experts (0.559 5), while attaining a high F1-score of 0.721 9. These findings have substantial significance for the long-term assessment of patients with IESS.

Reference | Supplementary Material | Related Articles | Metrics | Comments（0）

Select

Histological Image Diagnosis of Breast Cancer Based on Multi-Attention Convolution Neural Network

XU Wangwang^1，2 (徐旺旺), XU Liangfeng^1，2 (许良凤), LIU Ninghui³(刘宁徽), LU Na³(律娜)

J Shanghai Jiaotong Univ Sci 2025, 30 (1): 91-106. DOI: 10.1007/s12204-024-2705-4

Abstract （973）

PDF（pc）（1716KB）（208）

Save

Breast cancer is a serious and high morbidity disease in women, and it is the main cause of cancer death in China. However, getting tested and diagnosed early can reduce the risk of cancer. At present, there are clinical examinations, imaging screening and biopsies, among which histopathological examination is the gold standard. However, the process is complicated and time-consuming, and misdiagnosis may exist. This paper puts forward a classification framework based on deep learning, introducing multi-attention mechanism, selecting kernel convolution instead of ordinary convolution, and using different weights and combinations to pay attention to the accuracy index and growth rate of the model. In addition, we also compared the learning rate regulators. Error function can fine-tune the learning rate to achieve good performance, using label softening to reduce the loss error caused by model error recognition in the label, and assigning different category weights in the loss function to balance the positive and negative samples. We used the BreakHis data set to automatically classify histological images into benign and malignant, four categories and eight subtypes. Experimental results showed that the accuracy of binary classifications ranged from 98.23% to 99.50%, and that of multipl classifications ranged from 97.89% to 98.11%.

Reference | Related Articles | Metrics | Comments（0）

Select

Positional Information is a Strong Supervision for Volumetric Medical Image Segmentation

ZHAO Yinjie¹ (赵寅杰), HOU Runpingg¹ (侯润萍), ZENG Wanqin² (曾琬琴), QIN Yulei¹ (秦玉磊), SHEN Tianle² (沈天乐), XU Zhiyong² (徐志勇), FU Xiaolong^2* (傅小龙), SHEN Hongbin^1* (沈红斌)

J Shanghai Jiaotong Univ Sci 2025, 30 (1): 121-129. DOI: 10.1007/s12204-023-2614-y

Abstract （765）

PDF（pc）（482KB）（165）

Save

Medical image segmentation is a crucial preliminary step for a number of downstream diagnosis tasks. As deep convolutional neural networks successfully promote the development of computer vision, it is possible to make medical image segmentation a semi-automatic procedure by applying deep convolutional neural networks to finding the contours of regions of interest that are then revised by radiologists. However, supervised learning necessitates large annotated data, which are difficult to acquire especially for medical images. Self-supervised learning is able to take advantage of unlabeled data and provide good initialization to be finetuned for downstream tasks with limited annotations. Considering that most self-supervised learning especially contrastive learning methods are tailored to natural image classification and entail expensive GPU resources, we propose a novel and simple pretext-based self-supervised learning method that exploits the value of positional information in volumetric medical images. Specifically, we regard spatial coordinates as pseudo labels and pretrain the model by predicting positions of randomly sampled 2D slices in volumetric medical images. Experiments on four semantic segmentation datasets demonstrate the superiority of our method over other self-supervised learning methods in both semisupervised learning and transfer learning settings. Codes are available at https://github.com/alienzyj/PPos.

Reference | Related Articles | Metrics | Comments（0）

Select

Brain Age Detection of Alzheimer’s Disease Magnetic Resonance Images Based on Mutual Information - Support Vector Regression

LIU Yuchuan¹ (刘玉川), LI Hao¹ (李浩), TANG Yulong¹ (唐宇龙), LIANG Dujuan² (梁杜娟), TAN Jia³ (谭佳), FU Yue¹ (符玥), LI Yongming^4∗ (李勇明)

J Shanghai Jiaotong Univ Sci 2025, 30 (1): 130-135. DOI: 10.1007/s12204-023-2590-2

Abstract （988）

PDF（pc）（636KB）（155）

Save

Brain age is an effective biomarker for diagnosing Alzheimer’s disease (AD). Aimed at the issue that the existing brain age detection methods are inconsistent with the biological hypothesis that AD is the accelerated aging of the brain, a mutual information - support vector regression (MI-SVR) brain age prediction model is proposed. First, the age deviation is introduced according to the biological hypothesis of AD. Second, fitness function is designed based on mutual information criterion. Third, support vector regression and fitness function are used to obtain the predicted brain age and fitness value of the subjects, respectively. The optimal age deviation is obtained by maximizing the fitness value. Finally, the proposed method is compared with some existing brain age detection methods. Experimental results show that the brain age obtained by the proposed method has better separability, can better reflect the accelerated aging of AD, and is more helpful for improving the diagnostic accuracy of AD.

Reference | Related Articles | Metrics | Comments（0）

Select

Medical Image Encryption Based on Fisher-Yates Scrambling and Filter Diffusion

HUANG Jiaxin (黄佳鑫), GUO Yali (郭亚丽), GAO Ruoyun (高若云),LI Shanshan^∗ (李珊珊)

J Shanghai Jiaotong Univ Sci 2025, 30 (1): 136-152. DOI: 10.1007/s12204-023-2618-7

Abstract （1023）

PDF（pc）（8077KB）（183）

Save

A medical image encryption is proposed based on the Fisher-Yates scrambling, filter diffusion and S-box substitution. First, chaotic sequence associated with the plaintext is generated by logistic-sine-cosine system, which is used for the scrambling, substitution and diffusion processes. The three-dimensional Fisher-Yates scrambling, S-box substitution and diffusion are employed for the first round of encryption. The chaotic sequence is adopted for secondary encryption to scramble the ciphertext obtained in the first round. Then, three-dimensional filter is applied to diffusion for further useful information hiding. The key to the algorithm is generated by the combination of hash value of plaintext image and the input parameters. It improves resisting ability of plaintext attacks. The security analysis shows that the algorithm is effective and efficient. It can resist common attacks. In addition, the good diffusion effect shows that the scheme can solve the differential attacks encountered in the transmission of medical images and has positive implications for future research.

Reference | Related Articles | Metrics | Comments（0）

Select

Improvement of Prior Image for Metal Artifact Reduction of Computed Tomography

Sun Wenwu, Zhuang Tiange, Chen Siping

J Shanghai Jiaotong Univ Sci 2025, 30 (3): 446-454. DOI: 10.1007/s12204-023-2643-6

Abstract （488）

PDF（pc）（769KB）（165）

Save

It is not easy to reduce the metal artifacts of computed tomography images. However, the pixel values inside the metal artifact regions vary smoothly, while those on the borders of the metal and the bone regions vary sharply. When the Canny operation by adaptive thresholding is conducted on the raw image, the almost continuous edges can be formed obviously on the borders of the metal and the bone regions, but this kind of information cannot be formed for the metal artifact regions. In this paper, by searching the closed areas formed by the border edges of the bone regions in the Canny image, the metal artifact regions, which are very difficult to discriminate only by intensity thresholding, can be excluded effectively. A novel prior image-based method is thus developed for metal artifact reduction. The experiments demonstrate that the proposed method can be realized easily and reduce the metal artifacts effectively even if multiple large metal objects exist simultaneously in the image. The method is suitable for the clinical application.

Reference | Related Articles | Metrics | Comments（0）

Select

Real-Time Lightweight Convolutional Neural Network for Polyp Detection in Endoscope Images

Si Bingqi, Pang Chenxi, Wang Zhiwu, Jiang Pingping, Yan Guozheng

J Shanghai Jiaotong Univ Sci 2025, 30 (3): 521-534. DOI: 10.1007/s12204-023-2671-2

Abstract （686）

PDF（pc）（1563KB）（174）

Save

Colorectal cancer is the most common cancer with a second mortality rate. Polyp lesion is a precursor symptom of colorectal cancer. Detection and removal of polyps can effectively reduce the mortality of patients in the early period. However, mass images will be generated during an endoscopy, which will greatly increase the workload of doctors, and long-term mechanical screening of endoscopy images will also lead to a high misdiagnosis rate. Aiming at the problem that computer-aided diagnosis models deeply depend on the computational power in the polyp detection task, we propose a lightweight model, coordinate attention-YOLOv5-Lite-Prune, based on the YOLOv5 algorithm, which is different from state-of-the-art methods proposed by the existing research that applied object detection models or their variants directly to prediction task without any lightweight processing, such as faster region-based convolutional neural networks, YOLOv3, YOLOv4, and single shot multibox detector. The innovations of our model are as follows: First, the lightweight EfficientNetLite network is introduced as the new feature extraction network. Second, the depthwise separable convolution and its improved modules with different attention mechanisms are used to replace the standard convolution in the detection head structure. Then, the α-intersection over union loss function is applied to improve the precision and convergence speed of the model. Finally, the model size is compressed with a pruning algorithm. Our model effectively reduces parameter amount and computational complexity without significant accuracy loss. Therefore, the model can be successfully deployed on the embedded deep learning platform, and detect polyps with a speed above 30 frames per second, which means the model gets rid of the limitation that deep learning models must rely on high-performance servers.

Reference | Related Articles | Metrics | Comments（0）

Select

Image Mosaic Method of Capsule Endoscopy Intestinal Wall Based on Improved Weighted Fusion

Ma Ting, Wu Jianfang, Hu Feng, Nie Wei, Liu Youxin

J Shanghai Jiaotong Univ Sci 2025, 30 (3): 535-544. DOI: 10.1007/s12204-023-2637-4

Abstract （514）

PDF（pc）（2442KB）（171）

Save

There is still a dearth of systematic study on picture stitching techniques for the natural tubular structures of intestines, and traditional stitching techniques have a poor application to endoscopic images with deep scenes. In order to recreate the intestinal wall in two dimensions, a method is developed. The normalized Laplacian algorithm is used to enhance the image and transform it into polar coordinates according to the characteristics that intestinal images are not obvious and usually arranged in a circle, in order to extract the new image segments of the current image relative to the previous image. The improved weighted fusion algorithm is then used to sequentially splice the segment images. The experimental results demonstrate that the suggested approach can improve image clarity and minimize noise while maintaining the information content of intestinal images. In addition, the method’s seamless transition between the final portions of a panoramic image also demonstrates that the stitching trace has been removed.

Reference | Related Articles | Metrics | Comments（0）

Select

Computer Aided Diagnosis for COVID-19 in CT Images Utilizing Transfer Learning and Attention Mechanism

Fan Xinggang, Liu Jiaxian, Li Chao, Yang Youdong, Gu Wenting, Jiang Xinyang

J Shanghai Jiaotong Univ Sci 2025, 30 (3): 566-581. DOI: 10.1007/s12204-023-2646-3

Abstract （566）

PDF（pc）（911KB）（174）

Save

Various and intricate varieties of lung disease have made it challenging for computer aided diagnosis to appropriately segment lung lesions utilizing computed tomography (CT) images. This study integrates transfer learning with the attention mechanism to construct a deep learning model that can automatically detect new coronary pneumonia on lung CT images. In this study, using VGG16 pre-trained by ImageNet as the encoder, the decoder was established utilizing the U-Net structure. The attention module is incorporated during each concatenate procedure, permitting the model to concentrate on the critical information and identify the crucial components efficiently. The public COVID-19-CT-Seg-Benchmark dataset was utilized for experiments, and the highest scores for Dice, F1, and Accuracy were 0.907 1, 0.907 6, and 0.996 5, respectively. The generalization performance was assessed concurrently, with performance metrics including Dice, F1, and Accuracy over 0.8. The experimental findings indicate the feasibility of the segmentation network proposed in this study.

Reference | Related Articles | Metrics | Comments（0）

Select

CT-MFENet: Context Transformer and Multi-Scale Feature Extraction Network via Global-Local Features Fusion for Retinal Vessels Segmentation

Shao Dangguo, Yang Yuanbiao, Ma Lei, Yi Sanli

J Shanghai Jiaotong Univ Sci 2025, 30 (4): 668-682. DOI: 10.1007/s12204-024-2748-6

Abstract （471）

PDF（pc）（4077KB）（292）

Save

Segmentation of the retinal vessels in the fundus is crucial for diagnosing ocular diseases. Retinal vessel images often suffer from category imbalance and large scale variations. This ultimately results in incomplete vessel segmentation and poor continuity. In this study, we propose CT-MFENet to address the aforementioned issues. First, the use of context transformer (CT) allows for the integration of contextual feature information, which helps establish the connection between pixels and solve the problem of incomplete vessel continuity. Second, multi-scale dense residual networks are used instead of traditional CNN to address the issue of inadequate local feature extraction when the model encounters vessels at multiple scales. In the decoding stage, we introduce a local-global fusion module. It enhances the localization of vascular information and reduces the semantic gap between high- and low-level features. To address the class imbalance in retinal images, we propose a hybrid loss function that enhances the segmentation ability of the model for topological structures. We conducted experiments on the publicly available DRIVE, CHASEDB1, STARE, and IOSTAR datasets. The experimental results show that our CT-MFENet performs better than most existing methods, including the baseline U-Net.

Reference | Related Articles | Metrics | Comments（0）

Select

Ground-Glass Lung Nodules Recognition Based on CatBoost Feature Selection and Stacking Ensemble Learning

Miao Jun, Chang Yiru, Chen Chen, Zhang Maoyuan, Liu Yan, Qi Honggang, Guo Zhijun, Xu Qian

J Shanghai Jiaotong Univ Sci 2025, 30 (4): 790-799. DOI: 10.1007/s12204-024-2761-9

Abstract （416）

PDF（pc）（510KB）（240）

Save

Aimed at the issues of high feature dimensionality, excessive data redundancy, and low recognition accuracy of using single classifiers on ground-glass lung nodule recognition, a recognition method was proposed based on CatBoost feature selection and Stacking ensemble learning. First, the method uses a feature selection algorithm to filter important features and remove features with less impact, achieving the effect of data dimensionality reduction. Second, random forests classifier, decision trees, K-nearest neighbor classifier, and light gradient boosting machine were used as base classifiers, and support vector machine was used as meta classifier to fuse and construct the ensemble learning model. This measure increases the accuracy of the classification model while maintaining the diversity of the base classifiers. The experimental results show that the recognition accuracy of the proposed method reaches 94.375%. Compared to the random forest algorithm with the best performance among single classifiers, the accuracy of the proposed method is increased by 1.875%. Compared to the recent deep learning methods (ResNet+GBM+Attention and MVCSNet) on ground-glass pulmonary nodule recognition, the proposed method’s performance is also better or comparative. Experiments show that the proposed model can effectively select features and make recognition on ground-glass pulmonary nodules.

Reference | Related Articles | Metrics | Comments（0）

Select

Multi-Consistency Training for Semi-Supervised Medical Image Segmentation

Wu Changxue, Zhang Wenxi, Han Jiaozhi, Wang Hongyu

J Shanghai Jiaotong Univ Sci 2025, 30 (4): 800-814. DOI: 10.1007/s12204-024-2733-0

Abstract （387）

PDF（pc）（1126KB）（255）

Save

Medical image segmentation is a crucial task in clinical applications. However, obtaining labeled data for medical images is often challenging. This has led to the appeal of semi-supervised learning (SSL), a technique adept at leveraging a modest amount of labeled data. Nonetheless, most prevailing SSL segmentation methods for medical images either rely on the single consistency training method or directly fine-tune SSL methods designed for natural images. In this paper, we propose an innovative semi-supervised method called multi-consistency training (MCT) for medical image segmentation. Our approach transcends the constraints of prior methodologies by considering consistency from a dual perspective: output consistency across different up-sampling methods and output consistency of the same data within the same network under various perturbations to the intermediate features. We design distinct semi-supervised loss regression methods for these two types of consistencies. To enhance the application of our MCT model, we also develop a dedicated decoder as the core of our neural network. Thorough experiments were conducted on the polyp dataset and the dental dataset, rigorously compared against other SSL methods. Experimental results demonstrate the superiority of our approach, achieving higher segmentation accuracy. Moreover, comprehensive ablation studies and insightful discussion substantiate the efficacy of our approach in navigating the intricacies of medical image segmentation.

Reference | Related Articles | Metrics | Comments（0）

Topics