Deep Learning-Based CT Image Characteristics and Postoperative Anal Function Restoration for Patients with Complex Anal Fistula

Objective This study aimed to optimize the CT images of anal fistula patients using a convolutional neural network (CNN) algorithm to investigate the anal function recovery. Methods 57 patients with complex anal fistulas admitted to our hospital from January 2020 to February 2021 were selected as research subjects. Of them, CT images of 34 cases were processed using the deep learning neural network, defined as the experimental group, and the remaining unprocessed 23 cases were in the control group. Whether to process CT images depended on the patient's own wish. The imaging results were compared with the results observed during the surgery. Results It was found that, in the experimental group, the images were clearer, with DSC = 0.89, precision = 0.98, and recall = 0.87, indicating that the processing effects were good; that the CT imaging results in the experimental group were more consistent with those observed during the surgery, and the difference was notable (P < 0.05). Furthermore, the experimental group had lower RP (mmHg), AMCP (mmHg) scores, and postoperative recurrence rate, with notable differences noted (P < 0.05). Conclusion CT images processed by deep learning are clearer, leading to higher accuracy of preoperative diagnosis, which is suggested in clinics.


Introduction
Anal fistula, also known as "anal leakage," is a chronic inflammatory disease that starts with the anal gland and then invades the anal canal and other normal skin tissues around the anus [1], so it is also defined as inflammatory rectum disease (RD) [2]. e anal fistula is composed of the internal opening, the fistula, and the external opening [3]. Complex anal fistula refers to the one with 2 or more internal or external openings and fistulas [4], which will bring great inconvenience to the patient's life. Its incidence is reported to be approximately 0.01%, and the young are predominantly affected [5]. ere are two main treatment methods for anal fistula: one is conservative treatment based on traditional Chinese medicine, and the other is surgical treatment [6]. Nevertheless, for complex anal fistula, it is reported that the recurrence rate is as high as 30% to 50% or even 10% patients cannot be radically cured after another operation [7]. ere are many factors that affect surgical treatment. e main factor is whether the positioning of the internal mouth is accurate before surgery [8,9], and the symptomatic degree of the surgical plan will affect the probability of recurrence and the degree of anal function retention [10,11]. e complex anal fistula is mainly treated by the surgery, with traditional Chinese medicine used for adjuvant treatment. erefore, imaging examination is of great significance for the treatment of the complex anal fistula [12,13]. e commonly used imaging examination methods include CT, MRI, and B-ultrasound. However, in the examination of complex anal fistulas, CT has low resolution for soft tissues, such as pelvic tissue and anal sphincter, so the accuracy rate is only 24%∼60% [14]. With the rapid development of the technology, deep learning artificial intelligence technology is widely used in the field of imaging and the neural network system has made great progress in processing medical images, which has become a research hotspot [15]. Deep learning has solved two major problems in the medical field: one is the imbalance between the number of imaging doctors and clinical imaging data; the other is the imbalance in the level of imaging doctors and the allocation of resources [16]. Deep convolutional network models can learn image features from a large number of samples, achieving end-to-end classification and detection [17].
In this study, deep learning was used to process CT images of patients with complex anal fistula, which was expected to provide guidance for the clinical diagnosis and treatment of the disease.

Research
Subjects. 57 patients admitted to our hospital from January 2019 to February 2020 who had been diagnosed with complex anal fistula were selected as research subjects, including 35 males and 22 females, aged 18∼70, with an average age of (39.2 ± 2.1) years. All patients had the CT examination, and the CT images of 34 cases were processed using the convolutional neural network, defined as the experimental group, while the remaining unprocessed 23 cases were in the control group. Whether to process CT images depended on the patient's own wish. Clinically, all patients suffered from repeated swelling and pain around the anus, accompanied by pus discharge. e duration of the disease was 3 to 12 months, and the average duration of the disease was 7 months. Physical examination revealed purulent secretions around the anus and external opening on the skin surface. is study has been approved by the ethics committee, and the patients and their families have signed the informed consent forms. Intrarectal ultrasound, MRI, and CT examinations were conducted [18].
Patients were selected as per these inclusion criteria: (1) patients diagnosed with complex anal fistulas based on digital rectal examination, (2) the disease course lasted more than 3 months, (3) the length of the fistula was more than 3 cm (including 3 cm), (4) patients who had not undergone anal fistula surgery before, and (5) those aged between 18 and 75.
Exclusion criteria were as follows: (1) patients with the complex anal fistula caused by trauma, foreign body, infection, etc.; (2) patients accompanied by severe heart, lung, liver, and kidney dysfunction; (3) patients with cervical cancer, vaginal rectal fistula, rectal tumor, etc.; (4) patients in pregnancy and lactation; (5) patients with mental disorders; (6) patients with a history of pelvic and rectal radiotherapy; (7) patients with hematopoietic, endocrine, and immune system diseases.

CT Examination.
e anti-infection treatment was performed on all patients 1 to 2 weeks before the CT scan, to improve the symptoms of the patients to meet the examination standards. 2 hours before the scan, intestine cleaning was needed. e specific process is shown in Figure 1. All patients shared an instrument and the same skilled and experienced imaging doctor who was responsible for the CT examination and image reconstruction. After image reconstruction, 2 experienced radiologists and anorectal doctors read the films together and recorded the various conditions of the fistula. Surgical plans were formulated based on the results of the CT examination and other auxiliary examinations.

CT Image Optimization Model Based on Convolutional
Neural Network Algorithm. In this research, the convolutional neural network was used to optimize CT images. e convolutional neural network consists of the convolutional layer, the pooling layer, the fully connected layer, and the deconvolutional layer [19].
First, it is needed to collect the feature map of the CT image, provided that the convolutional layer has q layers, and then the feature map of the CT image input at the q-1 layer can be expressed by equation (1). en, the feature map of the CT image input at the q − 1 layer can be expressed by the following equation: where Y (q,p,b) is the convolutional kernel, b is the number of feature maps, and Y (q,p,b) is the bias. e activation function is a sigmoid transfer function, with a domain of (−∞, +∞) and a value domain of (0, 1). It is continuously derivable in the domain, and its reciprocal f'(x) is easy to calculate. e specific equation is as follows:

(2)
Assume T is the CT image and U represents all pixels in T. e segmentation part of T is represented by O. en, the probability that U n outputs O i in the i-th path can be expressed as follows: where α(O i ) is the value of Oi and A is a regularization term. en, the S n predicted by U n can be expressed as follows: e corresponding loss function is as follows: To minimize the loss function, the stochastic gradient descent [20] is used to solve equation (5).
If the function to be fitted is expressed as follows: then the square loss function is 2 Journal of Healthcare Engineering e stochastic gradient descent method is used to update δ j according to the sample: e network parameters are obtained by solving equation (9), so as to obtain the network model as follows: e processed pelvic CT image through the CNN algorithm is shown in Figure 2.

Evaluation of the Processed CT Image
2.3.1. Dice Similarity Coefficient. Dice similarity coefficient (DSC) is used to measure the similarity of two sets. e data pile can be regarded as a set, and the Dice distance is used to measure the similarity of strings. e DSC is expressed as follows: where G is the algorithm segmentation result and J is the gold standard. If the DSC is closer to 1, the segmentation result is more accurate. If the DSC value is infinitely close to 0, it indicates that the segmentation result basically does not overlap with the gold standard and the segmentation fails.

Precision.
Precision refers to the proportion of the number of correctly classified positive samples to the total number of classified positive samples. A closer precision value to 1 indicates a lower error rate. e precision is expressed as follows: where TP stands for true positive, which is consistent with the gold standard; on the contrary, FP stands for false positive, which is inconsistent with the gold standard. Journal of Healthcare Engineering rate that the positive samples are incorrectly classified as negative samples. Recall is expressed as follows: where TP stands for true positive, FN stands for false negative, and the classification is also inconsistent with the gold standard. Finally, it is concluded that DSC � 0.89, precision � 0.98, and recall � 0.87, which proves the feasibility of the model established in Section 2.3 and the effectiveness of image processing.

Observation Indicators
(1) e number and position of the internal and external openings of the fistula, whether the rectum, anal sphincter, and levator ani, are associated with the fistula, the area and length of the fistula, the diameter of the fistula, the fistula trend, whether there are complicated branches, and whether there is abscess formation. (2) e recovery of anal function of the two groups of patients after surgical treatment was observed based on the CT images before and after the treatment. e Wexner anal incontinence score (Table 1) and anorectal manometry were used to evaluate its function. e parameters of anorectal manometry were anal rest pressure (ARP) and anal maximal construction pressure (AMCP).

2.5.
Statistics. SPSS 22.0 was used to process data. Count data were expressed by percentage (%), the Y2 test was performed, and the rank sum test was performed for grade count data. e measurement data were expressed by the mean ± standard deviation, and the t-test was performed. P < 0.05 was the threshold for significance. Table 2, there was no notable difference in age, gender, and disease course of patients in the control group and the experimental group (P > 0.05), suggesting the feasibility of this study.

Examination Results and the Results after Surgical
Treatment. After analyses, there was no difference in the external opening, while parameters such as the internal openings, the trend of the fistula, whether there was branch, whether the fistula was associated with the perianal muscle, and fistula diameter ≥ 2 mm were different from the results observed during the surgical treatment in both the experimental group and the control group, and notable differences were noted between the CT examination results and the postoperative results in both groups (P < 0.05), suggesting that the accuracy of CT examination was insufficient (Table 3). Figure 3 shows the CT images of the two groups of patients. Obviously, the optimized CT images were clearer. Table 4 shows the consistency between the examination results and the postoperative results. It was observed that in the experimental group, the consistency of the CT examination results and the postoperative results were higher than in the control group (P < 0.05) (Figure 4).

Postoperative Anal Function Recovery and Postoperative
Recurrence Rate. Table 5 shows the statistics of Wexner anal incontinence score and anorectal manometry scores. It was found that the postoperative Wexner anal incontinence score of the experimental group was lower than that of the control group, showing notable differences. Additionally, the experimental group had lower RP (mmHg), AMCP (mmHg) scores, and the postoperative recurrence rate, and the difference was notable (P < 0.05). Figure 5 shows the CT images and the postoperative recovery of the two groups of patients, suggesting that both had good curative effects.

Discussion
Recently, CT imaging has been widely used in the examination of anal fistula diseases and it has certain advantages compared with other imaging techniques. Researchers have conducted a comparative study on the ordinary X-ray and CT imaging. e study found that the diagnosis of anal fistula with CT imaging is almost the same as the results of the surgery, with 96.7% of patients radically cured at one time, while the one-time cure rate of X-ray angiography is 70% [21]. e trend of fistulas is diverse and extremely irregular. Conventional two-dimensional CT images cannot          fully display its distribution, and it is impossible to make accurate judgment. Hence, it is urgent to optimize the technology. ere is abundant research combining CT imaging technology with three-dimensional reconstruction and applying them to the examination of anal fistulas. e results found that the accuracy, sensitivity, and specificity of the diagnosis of the internal opening of anal fistula have reached more than 90%, suggesting good effects [15]. e artificial intelligence in deep learning has made good progress in the field of medicine, especially in the field of medical imaging. Higaki et al. [22] applied fully convolutional neural networks and dense 3D conditional random field (CRF) technology to CT images for liver segmentation studies, to understand liver lesions. Tao et al. [23] applied FCN technology to the automatic analysis of cardiac magnetic resonance (CMR). In this study, the deep neural network was used to process the CT images of patients with anal fistulas. It was found that the deep learning-based CT imaging demonstrated good examination, diagnosis, and treatment results. Studies have also shown that higher accuracy of the preoperative diagnosis leads to better surgical effects, a greater possibility of recovery of anal function, and a lower probability of recurrence. Some experts have statistically studied 17 patients with complex anal fistula who underwent MSCT angiography diagnosis before the surgery. Some experts have statistically studied 17 patients with complex anal fistula who underwent multislice CT (MSCT) angiography diagnosis before the surgery. It was found that the consistency between the preoperative diagnosis and the surgical findings was close to 100%, and in the 1-year followup survey, there was no recurrence and the anal function recovered well [24]. In conclusion, deep learning-based CT imaging is worthy of promotion for the preoperative diagnosis of patients with anal fistula.

Conclusion
In this study, the deep learning neural network was used to process the CT images of patients with anal fistula and the preoperative imaging results were compared with the results observed during the surgery. It was found that in the experimental group, the CT imaging results were more consistent with those observed during the surgery (P < 0.05) and the images were clearer. Furthermore, the anal function recovery in the experimental was better.
To sum up, CT images processed by deep learning are clearer, leading to higher accuracy of preoperative diagnosis, better surgical effects, a higher possibility of recovery, and a lower recurrence rate, which is suggested in clinics.

Data Availability
No data were used to support this study.

Conflicts of Interest
e authors declare that they have no conflicts of interest.