Convolutional neural network to explore the effect of the drug on postoperative POCD in elderly patients with hip fracture

Abstract

The incidence of POCD may be further increased in elderly patients due to degenerative changes in the central nervous system. Once POCD occurs in these patients, it will not only prolong the length of hospital stay, increase medical expenses, but also seriously affect the quality of life of patients, delay the postoperative rehabilitation process, and bring a heavy burden to the family and society. Based on this, this study combines with image recognition technology to study the effect of trypsin inhibitor on postoperative POCD in elderly patients with hip fracture. The hip CT image segmentation algorithm based on concatenated convolutional neural network is used to realize the automatic phased segmentation of hip CT images. In addition, this study combines with image analysis to study the effect of trypsin inhibitor on postoperative POCD in elderly patients with hip fracture, and the image analysis method was based on the previous research methods. The research results show that the proposed method has certain effects.

Keywords

Trypsin elderly hip fracture postoperative cognitive function

1 Introduction

At present, the domestic automatic film production technology based on pathology has been very mature. Pathological sections are of great significance for the accurate diagnosis of the disease, and doctors generally analyze pathological images based on experience and semi-automated tools [1]. This type of manual detection requires a lot of doctor’s effort and relies on the doctor’s experience. Due to the subjective differences of doctors, the analysis process will inevitably lead to misdiagnosis.

With the rapid development of information technology and the maturity of computer image processing technology, computer diagnostic results are used to assist doctors in analyzing medical images, which can improve the accuracy of diagnosis and reduce the time for manual analysis of images. The goal of medical image segmentation is to segment the medical image into different anatomical structures and extract ROI regions, such as tumor tissue in the image. This basic demand has led to the emergence of a large number of segmentation algorithms, which reduces the labor of medical scientists while improving efficiency and accuracy, avoids the instability caused by manual operations, and has greatly promoted the development of biomedicine [2].

Image segmentation has been applied in many fields, including content-based image retrieval, machine vision, target detection, traffic control systems, and medical image analysis [3]. The medical image includes 3D images from X-rays and magnetic resonance in addition to 2D slices. Moreover, image segmentation is often used in the medical field to locate tumors and pathologies, confirm tissue volume, and diagnose and study anatomical structures. With the rapid development of information technology, a large number of new medical imaging technologies, such as X-ray, CT, and nuclear magnetic resonance, have appeared in the medical field. Doctors can quickly find the lesion through medical imaging and diagnose the condition as soon as possible. However, due to the high precision of medical image pixels and many interference factors, it is difficult for the human eye to distinguish between pathological tissue and normal tissue. Therefore, the introduction of image segmentation in the medical field has great significance for improving the accuracy of pathological diagnosis [4].

Medical decision support system (CDSS) is one of the main research directions in medical image processing [5]. Computer diagnostic results are used to assist doctors in analyzing medical images, which can improve the accuracy of diagnosis and reduce the time for manual analysis of images. Medical image segmentation is the basis of the CDS system, and the goal of segmentation is to segment the medical image into different anatomical structures and extract ROI regions, such as tumor tissue in the image. This basic requirement has led to the emergence of a large number of segmentation algorithms. However, the traditional segmentation algorithm basically relies on information such as edges, regions and shapes. Therefore, it relies on hand-extracted features and prior knowledge. Due to the cumbersome and unstable segmentation of artificial interaction, the widespread use of these algorithms is prevented.

The occurrence of POCD in elderly hip surgery weakens the patient’s response to various adverse factors and is more prone to complications such as hypostatic pneumonia, deep phlebitis of lower extremities, hemorrhoids, and indirectly affect the therapeutic effect of surgery. The early effects of POCD on the operation of senile ankle fractures are mainly: Limb activity in manic patients can cause fracture relocation, internal fixation fracture, slippage or joint dislocation; Depressed patients have reduced responsiveness to external stimuli and are complicated with complications such as hypo-pneumonia, hemorrhoids, and deep vein thrombosis of lower extremities. The late effects are mainly manifested in [6]: POCD will reduce the surgical outcome of patients and increase postoperative mortality. A large number of clinical studies and animal experiments have confirmed that the longer the duration of POCD, the worse the long-term recovery effect.

Postoperative cognitive dysfunction (POCD) is a common neurological complication after surgery. It is characterized by reversible damage in many aspects such as memory and mental concentration. Patients present with mental disorders, anxiety, personality changes, cognitive decline, and memory loss, which can last for months or years. Moreover, a small number of patients can develop permanent damage. According to reports in the literature, the incidence of POCD after non-cardiac surgery can be as high as 7 to 26%, the incidence of POCD at 6 days after joint replacement is as high as 72%, and there is still a 30% incidence after 6 months. Moreover, the incidence of POCD may be further increased in elderly patients due to degenerative changes in the central nervous system. Once POCD occurs in these patients, it will not only prolong the length of hospital stay, increase medical expenses, but also seriously affect the quality of life of patients, delay the postoperative rehabilitation process, and bring a heavy burden to the family and society. Based on this, this study combined with image recognition technology to study the effect of trypsin inhibitor on POCD after hip fracture in elderly patients [7].

2 Related work

In order to accurately segment the bone tissue from the microscopic images of the pathological section, the researchers have done a lot of work and proposed many effective segmentation algorithms. Currently, widely used algorithms include threshold segmentation, active contouring, edge detection, and morphology. At the same time, domestic and foreign scholars have proposed a variety of segmentation algorithms for medical images of specific applications [8], such as active contours, multispectral and other methods.

The techniques used in medical image segmentation vary with imaging methods and body parts. The requirement of brain image segmentation is different from the technique used for image segmentation of bone tissue. In the brain image, partial volume effect plays a major role. However, in bone tissue images, motion artifacts play a major role [9].

Unsupervised automatic medical image segmentation methods mainly include threshold [10], energy function [11] and region merging [12]. The threshold-based method extracts the ROI region according to a certain threshold, and the threshold is generally calculated by analyzing a predefined image feature such as a density histogram [13]. The energy function-based method obtains the ROI region boundary by minimizing the loss function. The loss function is usually defined according to the shape characteristics of the image and the statistical distribution threat. The method of region merging recursively combines different pixels [14]. Threshold-based methods are limited by the density distribution of the ROI region. If there are multiple extreme points in the density distribution, the segmentation may fail.

The shape-based correlation method uses a priori knowledge (shape and appearance information) about the ROI region to build the model, and then the model is applied to the image to segment the ROI region [15]. A method based on a trained classifier first extracts pixel or region features (SIFT, HOG, texture features) and then extracts the ROI region from the background image using various classifiers, such as support vector machines. Model-based methods are generally used in ROI regions with strong shape prior knowledge and fixed positions. This makes it impossible to apply to areas of ROI that have different sizes and random locations, such as tumors spread into different tissues, fractures, and infarcts. The above supervised method relies on low-level features and cannot obtain image level changes. The final segmentation result depends on whether the parameters are properly adjusted and effective preprocessing techniques.

The well-known supervised automatic segmentation method based on deep learning mainly includes a network using a pixel patch for training and an FCN network. A patch-based method predicts whether the center of a block of pixels or a block of pixels is in the ROI region, and all blocks of pixels are trained and predicted in the deep neural network. Since accurate image segmentation requires prediction of each pixel in the image, the patch-based method is generally not efficient. Moreover, in the patch-based method, the process of training and prediction is independent, the spatial structure of the image will be missing, and the final segmentation result lacks coherence [16]. In the literature [17], the contour features and FCN are combined to limit the boundaries of the ROI region, resulting in more accurate segmentation results. In the literature [18], U-net is proposed for image segmentation. They modify the FCN structure and combine the feature maps generated by different network layers to obtain the final segmentation result. Other research work includes conditional random field CRF, graph segmentation and level sets. However, these methods have a poor ability to improve the segmentation results. Compared with the supervised network FCN, the optimization work is generally based on unsupervised learning or prior knowledge and uses low-level features.

3 Research method

The 3D spatial background information of the voxels can more accurately represent the detailed features of the CT image. Since training samples in medical image analysis are usually finite and small, efficient propagation of 3D image information is very important for training FCN. In order to improve the learning ability of FCN, this paper adopts a hybrid residual connection learning architecture to extend the data connection from the jump connection of a single residual unit to the whole hip coarse segmentation mixed residual full convolution neural network. Moreover, this article refers to this type of data connection as a long connection, which integrates long connections into the FCN, which can help enhance the transmission of the whole and some details. In this way, more feature performance can be obtained in limited training data. Moreover, a module having a short connection will be referred to as a “Resblock module” hereinafter.

First, the CT image extracts the feature value by a down sampling convolution layer having a convolution kernel size of 2×2×2 and a step size of 2. It is then activated by the activation function for the active layer of PReLU and transferred to the Resblock module with short connections for feature extraction. “Conv-PReLU-Resblock” was performed 5 times in the down sampling process to obtain features, and finally 256 16×16×4 down sampling feature maps were obtained. The number of convolution kernels used in each convolutional layer is 16, 32, 64, 128, and 256, respectively.

The down sampling image is extracted by a deconvolution layer having a convolution kernel size of 2×2×2 and a step size of 2 to obtain a deconvolution image. Then, the feature maps obtained by the corresponding Resblock in the long connection and the down sampling process are spliced together, and the spliced feature maps are input into the Resblock module to continue extracting features. In order to get more image details, this paper deconvolves the feature map obtained by Resblock to get more image details. Such “Sum-Resblock-UpConv” is performed a total of 5 times in the down sampling process, and a coarse segmentation result map is obtained. The number of convolution kernels used by each convolution layer is 256, 128, 64, 32, and 16, respectively.

The network prediction consisting of two volumes having the same resolution as the original input data is processed by the soft-max layer. The soft-max layer outputs the probability that each voxel belongs to the hip region and the background. This paper uses a new objective function based on the dice coefficient (Dice), defined as: $D = \frac{2 \sum_{i}^{N} p_{i} g_{i}}{\sum_{i}^{N} p_{i}^{2} + \sum_{i}^{N} g_{i}^{2}}$ (1)

In the formula, N is the total number of voxels, p_i is the network predicted hip bone volume, and g_i is the gold standard volume. The gradient produced by the differentiation of the above formula is: $\frac{\partial D}{\partial p_{i}} = 2 [\frac{g_{i} (\sum_{i}^{N} p_{i}^{2} + \sum_{i}^{N} g_{i}^{2}) - \sum_{i}^{N} p_{i} g_{i}}{(\sum_{i}^{N} p_{i}^{2} + \sum_{i}^{N} g_{i}^{2})}]$ (2)

Using this formula, we don’t need to assign weights to different categories of samples to establish the right balance between foreground voxels and background voxels. Finally, the binary map of the hip and non-hip bone regions is output.

The method in this paper was evaluated on the public CT dataset provided by the CSI 2014 Hip Split Challenge. The data set is publicly available and is open to the public on the Spine website. The data set is a spinal CT obtained during routine clinical routine work. The CT scan covers the entire thoracic and lumbar spine and does not require venography. The data has a resolution of 512×512 and a slice thickness of 1 mm. Each hip bone is manually segmented by an expert and assigned a different label, and the first hip bone is labeled 100, the second hip bone is labeled 200, and so on. The data set consists of 10 training sets and 10 test sets and contains a healthy hip bone dataset and a diseased hip bone dataset. In this paper, the Dice similarity coefficient (DSC) and the Average symmetrical sur-face distance (ASD) are used to estimate the segmentation accuracy. DSC and ASD are calculated by the following formulas: $DSC = 2 \frac{| V_{r} \cap V_{s} |}{| V_{r} | + | V_{s} |}$ (3) $ASD = \frac{1}{| S_{S} |} \sum_{i = 1}^{| S_{S} |} | | d_{i} (S_{S}, S_{r}) | |$ (4)

In the formula, V_r is the reference volume, V_s is the segmentation volume, S_r is the reference plane, S_S is the segmentation plane, and d_i is the minimum distance from the point on S_S to S. Since the pretreatment methods studied in this paper affect the final segmentation experiment results, the data used for training, testing, and comparison in this section are raw data that has not been preprocessed. For comparison, the hip bone was divided into three parts for comparative analysis. In the formula, T1-T12 is the thoracic vertebra and L1-L5 is the lumbar vertebra. It can be seen from the table that the average DSC of the segmentation of the mixed residual network is 92.2%, and the average ASD is 1.32 mm. The segmentation results of the coarse-divided mixed residual network are shown in Figs. 1 and 2.

Fig. 1

Rough identification image 1 of hip fracture.

Fig. 2

Rough identification image 2 of hip fracture.

The rough segmentation method is also compared with advanced V-Net and 3D U-Net. The comparison results are shown in Figs. 3 and 4. Since the down sampling operation of V-net and U-net causes the input image to be highly compressed resulting in information loss, it is difficult to recover the complete image information during the up sampling process. Although U-net (residual network) and V-net (mixed residual) aggregate lower-level voxel feature maps to higher layers to compensate for the losses caused by pooling operations, the structural details of splitting bones are still extremely challenging. The hybrid residual full convolutional neural network studied in this paper can make full use of the spatial details of 3D CT and can perform segmentation tasks accurately. The average segmentation time of each set of data in this method is 8 minutes. On the basis of improving the segmentation accuracy, the segmentation time is still acceptable. Compared with V-net and U-net, the hybrid residual full convolution neural network coarse segmentation method studied in this paper has higher precision.

Fig. 3

Segmentation DSC.

Fig. 4

Training loss.

Based on the VGG-16 network model, the segmented segmentation convolutional neural network in this paper has been adjusted accordingly. In this paper, the segmentation network is reduced by several convolutional layers compared to VGG-16, and the BN layer is added for batch normalization to speed up network convergence. The specific network structure is shown in Fig. 5.

Fig. 5

Fine- segmented segmentation convolutional neural network structure.

The parameters of each layer of the fine-segmented segmentation convolutional neural network are detailed as shown in Table 1. The soft-max layer in the segmented convolutional neural network is a multi-classifier layer, which is used to refine the segmentation and segmented mark. of the hip bone in the rough segmentation result graph of the mixed residual full convolutional neural network. The training set label of the fine-segmented deep convolutional network CNN is labeled as C ={ (I_n, L_n) , n = 0.1 }. In the formula, I_n is the original spine CT image, and L_n is the manually segmented image of the hip bone with the label. Each L_n contains 18 basic fact category labels k, and the category label k consists of 17 hip-shaped hand-divided images and the background of each voxel in the original spine CT image. Among them, n is the number of voxels; the loss function of the fine-segmented deep convolutional network CNN is: $L = \frac{- 1}{n} \sum_{i = 1}^{k} λ_{i} \times [\sum_{x \forall N_{i}} (\hat{p_{k}} (x))]$ (5) In the formula, $\hat{p_{k}}$ is the soft-max layer output category probability, i is the weight factor, n is the total number of voxels, N_i is a category;Moreover, k ∈ [0, 1, 2, …, 17] represents the correct basic fact category label, and x is the voxel point. The final multi-hip bone segmentation results with hip segment markers were obtained by refining the segmented segmentation convolutional neural network.

Table 1

Fine-segmented segmentation convolutional neural network parameters

Layer	Kernel	Stride	Pad	Outputs
Data	–	–	–	32×32×32
Conv1	3×3×3	1	0	32×32×32
Maxpool1	2×2×2	2	0	16×16×16
Conv2	3×3×3	1	0	16×16×16
Maxpool2	2×2×2	2	0	8×8×8
Conv3	3×3×3	1	0	8×8×8
Maxpool3	2×2×2	1	0	8×8×8
Conv4	3×3×3	1	0	8×8×8
Maxpool3	2×2×2	2	0	4×4×4
FC1		FC-4096
FC2		FC-1024
Soft-max		Soft-max lables = 18

4 Experimental study

This study combined with image analysis to study the effect of trypsin inhibitor on POCD in elderly patients with hip fracture. The image analysis method was based on the previous research methods. Sixty elderly patients who underwent surgery for lower extremity arthroplasty from August 2017 to August 2019 were selected. According to the computer random number, the pairing design was performed according to gender and age to determine the patient group. The experiment was double-blind. Except for the investigators, subjects, surgeons, and serum biochemical indicators were not aware of the grouping of patients.

Inclusion criteria: (1) Patients undergoing elective femoral head replacement, total hip replacement, or knee replacement; (2) Age≥65 years old; (3) ASA grade I∼II; (4) The subject has signed an informed consent form.

General anesthesia: Midazolam 1–1.5 mg, propofol 1–1.5 mg/kg, fentanyl 0.04 ug/kg, vecuronium 0.1 mg/kg were sequentially injected intravenously to induce intubation. After successful intubation, the breathing was controlled using a ventilator (Dräger, Fabius GS, Germany). The tidal volume is set to 8 to 10 ml/kg, the respiratory rate is 10 times/min, and 100% pure oxygen is inhaled, and the flow rate is 1 L/min. ETCO2 is around 40 mmHg and 1–1.5 MAC sevoflurane (Sevoflurane) is maintained. Anesthesia was maintained by continuous pumping of remifentanil, and vecuronium 0.03 mg/kg was given intermittently as needed.

Combined spinal and epidural anesthesia (Camel brand lumbar hard anesthesia package): The patient took the left lateral position. After the patient took L_2 - 3 or L_3 - 4 intervertebral space for epidural puncture successfully, the lumbar puncture needle was place. After the cerebrospinal fluid was observed to be clear and the reflux was smooth, 10 mg of bupivacaine (about 20 seconds) was slowly injected. Thereafter, the epidural catheter was placed 3-4 cm via an epidural needle and properly fixed. After the patient is supine, the absolute plane of the patient’s anesthesia is adjusted to be below T10. At an interval of 60 minutes, 3–5 ml of 0.75% ropivacaine (Nexus 750 mg 10 ml) was added via an epidural catheter.

During the operation, the patient’s ECG, blood pressure, heart rate and pulse oximetry (Philips G60 monitor) need to be continuously monitored. Moreover, it is necessary to record the anesthesia time, the operation time, the amount of various anesthetic drugs, the amount of intraoperative blood loss, the length of hospital stay, and postoperative adverse reactions.

Postoperative cognitive function analysis: In the preoperative, immediately after surgery (T0), 15 minutes after surgery (T15), 40 minutes (T40) after surgery, 1 day after surgery (D1), 3 days (D3), 7 days (D7), and 1 month after surgery (M1), the Post-operative Quality Recovery Scale (PQRS) was used to evaluate the recovery of physiology, adverse feelings, emotions, daily activities, cognitive functions, and daily living abilities. In the preoperative, 7 days after surgery (D7) and 1 month after surgery (M1), the Mini-Mental State Examination (MMSE) was used to evaluate cognitive function and compare it with the results of the PQRS scale (The score decreased by 1 standard deviation or more before surgery, and the cognitive function decreased, the diagnosis was POCD).

Measuring bone density accuracy: It is expressed using a control coefficient of variation of 1.0 CV% (Coefficient of variation). The randomly confide standardized phantoms are scanned before the daily test, and the phantom measurements are controlled to fluctuate within the horizontal line of the QC chart during operation. The standard phantom was used to correct the performance of the instrument before each test. All bone density detection operations were performed by the same technician. The operating technician received a training certificate from the International Society for Clinical Densitometry. The cumulative accuracy of the bone densitometer in our hospital reached: 0.242%. Figure 6 shows the geometry of the hip joint based on the scanned image.

Fig. 6

Geometrical parameters of the hip joint obtained based on the scanned image.

5 Research result

Postoperative psychology scale tests were completed in both groups. The PQRS scores of the UTI group were higher than those of the CON group at 15 min after surgery, 40 min after surgery, 1 d after surgery, 3 d after surgery, and 7 d after surgery. The difference between the two groups was statistically significant (P < 0.05; Table 2). However, there was no significant difference in the PQRS scores between the two groups (P > 0.05; Table 2).

Table 2
Comparison of PQRS scores between the two groups

Project UTI group CON group P

Before anesthesia 58.56±3.42 58.30±6.28 0.861

15 minutes after 45.254±4.68 33.87±11.57 0.000*

surgery

40 minutes after 47.36±4.23 38.22±7.75 0.000*

surgery

1d after surgery 38.44±2.77 33.59±5.73 0.000***

3d after surgery 32.84±3.45 29.91±5.37 0.029*

7d after surgery 36.96±3.22 33.95±6.47 0.046*

1 m after surgery 55.56±5.12 53.09±5.59 0.118

Project	UTI group	CON group	P
Before anesthesia	58.56±3.42	58.30±6.28	0.861
15 minutes after	45.254±4.68	33.87±11.57	0.000***
surgery
40 minutes after	47.36±4.23	38.22±7.75	0.000***
surgery
1d after surgery	38.44±2.77	33.59±5.73	0.000***
3d after surgery	32.84±3.45	29.91±5.37	0.029*
7d after surgery	36.96±3.22	33.95±6.47	0.046*
1 m after surgery	55.56±5.12	53.09±5.59	0.118

The MMSE scores of the two groups were lower than the preoperative MMSE scores at 7d and 1 m after operation. The MMSE scores of the UTI group at 7d and 1 m after operation were higher than those of the CON group. Moreover, the difference between the two groups was statistically significant (P < 0.05, Table 3).

Table 3

Comparison of MMSE scores between the two groups

Project	UTI group	CON group	P
Preoperative	29.62±2.19	29.17±2.15	0.27
7d after surgery	26.43±4.18	25.35±2.21	<0.0001
1 m after surgery	27.29±3.26	26.24±3.38	<0.0001

There was no significant difference in the incidence of POCD between the UTI group (20%, 16%) and the CON group (30.4%, 26.1%) at 7 days after surgery and 1 m after surgery (P > 0.05, Table 4).

Table 4

Comparison of POCD incidence rates between the two groups

Project	UTI group	CON group	P
7d after surgery	20% (5)	30.4% (7)	0.43
1 m after surgery	16% (4)	26.1% (6)	0.39

6 Analysis and discussion

Ulinastatin is mainly administered intravenously in the clinic, with a bioavailability of 100% and a drug half-life of about 40 min. After administration, the active ingredients rapidly accumulate in the kidney and liver, and the blood concentration reaches a peak in about 5 minutes, and the metabolic pathway is mainly metabolized by the kidney. The proportion of drugs excreted in urine and feces after 12 hours of administration was 73% and 2.3%, respectively. Previous clinical and basic studies have shown that ulinastatin, a urinary trypsin inhibitor, inhibits the activity of serine proteases and many other proteases. Moreover, it can reduce the fluidity of the biolipid membrane and interfere with the binding of the effector and the receptor. In addition, it can prevent macrophage release of pro-inflammatory factors and reduce excessive inflammation after surgery. Currently, it has been widely used in patients with acute pancreatitis and pancreatic surgery. In addition to inhibiting various digestive enzymes, correcting cellular metabolic disorders and improving hemodynamic parameters and avoiding systemic inflammatory response syndrome are also important mechanisms for the treatment of acute pancreatitis with ulinastatin. Ulinastatin also improves microcirculation perfusion during shock, inhibits multiple protease activities, stabilizes lysosomal membranes, and inhibits the production of myocardial depressant factor (MDF). Moreover, it can scavenge free radicals, inhibit superoxide production, help control the malignant progression of shock, reduce the occurrence of MODS, and better help the treatment of shock. In this study, the choice of anesthesia was not limited. The reason is that there is no significant difference in the effect of general anesthesia or spinal anesthesia on postoperative cognitive function in elderly patients. Some scholars have analyzed the literature on the effects of general anesthesia and spinal anesthesia on POCD. It was found that 23 out of 24 studies considered that the effects of spinal anesthesia and general anesthesia on POCD were similar. Campbell et al. concluded that there was no significant difference in cognitive ability between general anesthesia and local anesthesia.

In summary, the application of ulinastatin during the perioperative period can effectively improve postoperative cognitive function and promote postoperative recovery. Moreover, its role may be achieved by reducing the release of pro-inflammatory factors and increasing the secretion of anti-inflammatory factors to inhibit post-operative inflammatory responses.

7 Conclusion

In this paper, the hip segmentation CT image based on the hybrid residual full convolutional neural network is designed. The hip segmentation in the image is initially rough segmented, and the hip bone is segmented from the CT image to obtain preliminary results. It can be seen from a large number of experiments that the mixed residual convolutional neural network of the hip bone segmentation in this paper can quickly and accurately segment the hip bone region, providing a large amount of prior knowledge of the hip bone position and region for the next step of refinement segmentation. Secondly, this study combines with image analysis to study the effect of trypsin inhibitor on POCD in elderly patients with hip fracture. The image analysis method is based on the previous research methods. The results of the study indicate that the use of trypsin inhibitors in elderly patients with hip fracture can effectively improve postoperative cognitive function and promote postoperative recovery. Moreover, its role may be achieved by reducing the release of pro-inflammatory factors and increasing the secretion of anti-inflammatory factors to inhibit post-operative inflammatory responses. In a limited time, the research of this paper has achieved the expected effect. The hip CT image segmentation algorithm based on concatenated convolutional neural network is used to realize the automatic phased segmentation of hip CT images. However, the algorithms studied in this paper still need to be improved, such as optimizing the network structure, extending the network to other medical image segmentation, etc., which provides theoretical ideas and techniques for follow-up medical research and treatment.

References

Hooper

and Vellas

, Commentary: Fatty acids and Alzheimer’s disease: evidence on cognition and cortical β-amyloid from secondary analyses of the Multidomain Alzheimer Preventive Trial[J], Journal of Prevention of Alzheimers Disease 5(3) (2018), 168.

Hsu

J.L.

, Hsu

W.C.

, Chang

C.C.

, et al., Everyday cognition scales are related to cognitive function in the early stage of probable Alzheimer’s disease and FDG-PET findings[J], Sci Rep 7(1), 2017.

Clare O’Donoghue

, Murphy

S.E.

, Zamboni

, et al., APOE genotype and cognition in healthy individuals at risk of Alzheimer’s disease: A review[J], Cortex (2018), 104.

Davis

, O’Connell

, Johnson

, et al., Estimating Alzheimer’s Disease Progression Rates from Normal Cognition Through Mild Cognitive Impairment and Stages of Dementia[J], Current Alzheimer Research 2018.

Inferior parietal transcranial direct current stimulation with training improves cognition in anomic Alzheimer’s disease and frontotemporal dementia:[J], Alzheimers & Dementia Translational Research & Clinical Interventions 3(2) (2017), 247–253.

Portbury

S.D.

, Hare

D.J.

, Sgambelloni

, et al., Trehalose Improves Cognition in the Transgenic Tg2576 Mouse Model of Alzheimer’s Disease[J], Journal of Alzheimer’s disease: JAD 60(Suppl 1) (2017), 1–12.

Martínez-Sánchez

, Meilán

J.J.G.

, AV F.J., et al., Speech rhythm alterations in Spanish-speaking individuals with Alzheimer’s disease[J], Aging Neuropsychology and Cognition 24(4) (2017), 418–434.

Decker

J.D.

and Michael

, Pharmacological Treatment of Alzheimer’s Disease[J]. 2017.

, Qiang

, Li

, et al., An efficient biosensor for monitoring Alzheimer’s disease risk factors: modulation and disaggregation of the Aβ aggregation process[J], Journal of Materials Chemistry B (2019), 7.

10.

Mullai

, Sangeetha

, Surya

, Madhan kumar

, Jeyabalan

and Broumi

, A Single Valued Neutrosophic Inventory Model with Neutrosophic Random Variable, International Journal of Neutrosophic Science 1(2) (2020), 52–63.

11.

Zaied

A.N.H.

, Ismail

and El-Sayed

, A Survey on Meta-heuristic Algorithms for Global Optimization Problems, Journal of Intelligent Systems and Internet of Things 1(1) (2020), 40–47.

12.

Gadicha

and Gadicha

, Implicit Authentication Approach by Generating Strong Password through Visual Key Cryptography, Journal of Cybersecurity and Information Management 1(1) (2020), 5–16.

13.

Wenfei

, Penglin

, Chuanghe

, et al., Progress in research on association of cognitive reserve with disease-related cognitive disorders[J], Journal of Neurology & Neurorehabilitation, 2017.

14.

, Piers

R.J.

and Devine

, How technology is reshaping cognitive assessment: Lessons from the Framingham Heart Study[J], Neuropsychology 31(8) (2017), 846–861.

15.

A clinical dose of angiotensin-converting enzyme (ACE) inhibitor and heterozygous ACE deletion exacerbate Alzheimer’s disease pathology in mice[J], Journal of Biological Chemistry, 2019.

16.

Cammisuli

D.M.

, Innocenti

, Fusi

, et al., Aerobic exercise effects upon cognition in Alzheimer’s Disease: A systematic review of randomized controlled trials[J], Archives Italiennes de Biologie 156(1-2) (2018), 54–63.

17.

Perkovic

M.N.

, Strac

D.S.

, Tudor

, et al., Catechol-O-methyltransferase, cognition and Alzheimer’s disease[J], Current Alzheimer Research (2017), 15.

18.

Wei

E.X.

, Oh

E.S.

, Harun

, et al., Vestibular Loss Predicts Poorer Spatial Cognition in Patients with Alzheimer’s Disease[J], Journal of Alzheimer’s disease: JAD 61(3) (2017), 1–9.