Electronic device and method for classifying small bowel obstruction by using abdominal computed tomography image

Abstract

An electronic device for classifying small bowel obstruction by using an abdominal computed tomography (CT) image, according to one embodiment of the present invention, comprises a processor for acquiring classification information by inputting an abdominal CT image into a deep learning model trained to classify either high-grade small bowel obstruction (HGSBO) or low-grade small bowel obstruction (LGSBO) from the abdominal CT image, wherein the deep learning model includes: a feature extractor for extracting feature information including CT depth information from the abdominal CT image; a first classifier for classifying the abdominal CT image to be either normal or abnormal by using the feature information; and a second classifier for classifying the abdominal CT image as either HGSBO or LGSBO by using the feature information, if the abdominal CT image is abnormal.

Description

Electronic device and method for classifying small bowel obstruction using abdominal computed tomography images

[0001] The present invention relates to an electronic device and method for classifying small bowel obstruction using abdominal computed tomography images.

[0002] Acute small bowel obstruction (ASBO) is a common acute abdominal condition that may require urgent surgery. Delayed surgery for acute small bowel obstruction is known to have a significant morbidity and mortality rate. The most common cause of acute intestinal obstruction is postoperative adhesions (70%), followed by cancer, inflammatory bowel disease, and hernias. The pathophysiology of acute small bowel obstruction begins with the onset of intestinal obstruction due to various causes and progresses to the worsening of proximal intestinal dilatation due to the accumulation of fermented gas and displaced fluids, leading to the obstruction. Prolonged intestinal obstruction and distension can lead to hypovolemia and increased intraintestinal pressure, which can lead to intestinal ischemia or necrosis. Ischemia and intestinal wall dilatation weaken the intestinal barrier, increasing the risk of intestinal perforation. This risk increases with clinical deterioration and duration of treatment failure.

[0003] Treatment options for acute small bowel obstruction vary significantly depending on the pathophysiological progression. If intestinal infarction or peritonitis is suspected, emergency surgery is essential. Furthermore, if intestinal ischemia worsens, early surgery is necessary. However, non-operative management (NOM) is feasible in approximately 70% of cases, and approximately 20% of patients who undergo NOM ultimately undergo surgery due to clinical deterioration.

[0004] Recently, abdominal computed tomography (CT) has been shown to play an important role in predicting NOM treatment failure, and several studies have reported the use of abdominal CT to perform imaging findings and modeling (Rondenet C, Millet I, Corno L, et al. CT diagnosis of closed-loop bowel obstruction mechanism is not sufficient to indicate emergent surgery. EurRadiol 2020;30:1105-12. etc.). However, although abdominal CT plays an important role in deciding on surgical treatment for acute small bowel obstruction, radiological findings requiring surgery are detected only by experienced radiologists, and the inter-observation variability is relatively high.

[0005] Artificial intelligence (AI) has recently been widely applied to various tasks in medical imaging. Most studies on acute small bowel obstruction and AI have been conducted using simple radiography. Furthermore, AI studies using abdominal CT are rare, and recent studies have relied on human interpretation rather than AI-based approaches.

[0006] Although the use of AI in abdominal CT has been reported in the literature, it still contains several issues. For example, a recent study by Vanderbecq et al. successfully detected the transition zone of acute small bowel obstruction with CT, but only used CT images of acute small bowel obstruction and not normal CT images (Vanderbecq Q, Ardon R, De Reviers A, et al. Adhesion-related small bowel obstruction: deep learning for automatic transition-zone detection by CT. Insights Imaging 2022;13:13). Although the proposed method diagnosed several cases of acute small bowel obstruction on CT, it has limitations in that it cannot classify normal and abnormal cases because it only used abnormal data. Furthermore, a pilot study was previously conducted to determine whether AI could distinguish between normal and acute small bowel obstruction X-ray images, but the imbalance between normal and abnormal data was significant and non-CT images were used.

[0007] Thus, diagnosing patients at risk for requiring emergency surgery is often a clinically challenging situation. Furthermore, distinguishing between high-risk cases, such as closed-loop obstructions or small bowel obstructions due to banding, presents a challenging and challenging task for radiologists and clinicians.

[0008] The purpose of the present invention is to provide a deep learning model that can effectively diagnose acute small bowel obstruction and support early diagnosis and identification of patients at risk for surgery.

[0009] In one embodiment of the present invention, an electronic device for classifying small bowel obstruction using an abdominal computed tomography (CT) image includes a processor for obtaining classification information by inputting an abdominal CT image of a subject into a deep learning model trained to classify the abdominal CT image into any one of normal, high-grade small bowel obstruction (HGSBO), and low-grade small bowel obstruction (LGSBO), and the like, wherein the deep learning model may be configured to: a feature extractor for extracting feature information including CT depth information from the abdominal CT image; a first classifier for classifying the abdominal CT image into any one of normal and abnormal using the feature information; and a second classifier for classifying the abdominal CT image into any one of high-risk small bowel obstruction and low-risk small bowel obstruction using the feature information in the case of abnormality.

[0010] The above processor can obtain feature information including the CT depth information by performing pooling that preserves depth information for the vertical and horizontal spaces of the abdominal CT image through the feature extractor.

[0011] The above processor can train a first classifier to classify an abdominal CT image as either normal or abnormal using feature information of a normal abdominal CT image and an abnormal abdominal CT image.

[0012] The above processor can train a second classifier to classify the abdominal CT image into either a high-risk small bowel obstruction or a low-risk small bowel obstruction using feature information and CT depth information of the abnormal abdominal CT image.

[0013] In one embodiment of the present invention, a method for classifying small bowel obstruction using an abdominal computed tomography (CT) image performed by an electronic device includes the step of obtaining classification information by inputting an abdominal CT image of a subject into a deep learning model trained to classify the abdominal CT image into any one of normal, high-grade small bowel obstruction (HGSBO), and low-grade small bowel obstruction (LGSBO), wherein the deep learning model may be configured to: a feature extractor for extracting feature information including CT depth information from the abdominal CT image; a first classifier for classifying the abdominal CT image into any one of normal and abnormal using the feature information; and a second classifier for classifying the abdominal CT image into any one of high-risk small bowel obstruction and low-risk small bowel obstruction using the feature information in the case of abnormality.

[0014] The step of obtaining the above classification information may include a step of obtaining feature information including the CT depth information by performing pooling that preserves depth information for the vertical and horizontal spaces of the abdominal CT image through the feature extractor.

[0015] The method may further include a step of training a first classifier to classify an abdominal CT image as either normal or abnormal using feature information of a normal abdominal CT image and an abnormal abdominal CT image.

[0016] The method may further include a step of training a second classifier to classify the abdominal CT image into either a high-risk small bowel obstruction or a low-risk small bowel obstruction using feature information and CT depth information of the abnormal abdominal CT image.

[0017] According to one embodiment of the present invention, an artificial intelligence-based deep learning model can be constructed to help detect high-risk patients with acute small bowel obstruction (ASBO) using abdominal CT images.

[0018] According to one embodiment of the present invention, a deep learning classifier can be used to effectively detect high-risk small bowel obstruction patients using only abdominal CT images.

[0019] A novel 3D classifier with a dual-branching structure and depth-preserving pooling based on the artificial intelligence algorithm of the present invention can be a reliable screening and decision support tool for high-risk patients with small bowel obstruction.

[0020] FIG. 1 is a schematic diagram illustrating the operation of an electronic device according to one embodiment of the present invention.

[0021] FIG. 2 is a block diagram illustrating the configuration of an electronic device according to one embodiment of the present invention.

[0022] FIG. 3 is a diagram illustrating the structure of a deep learning model according to one embodiment of the present invention.

[0023] FIG. 4 is a diagram illustrating an operation flow diagram of an electronic device according to one embodiment of the present invention.

[0024] FIG. 5 is a diagram illustrating a heat map using Grad-CAM of a deep learning model according to one embodiment of the present invention and an existing deep learning model.

[0025] Hereinafter, preferred embodiments of the present invention will be described in detail with reference to the accompanying drawings. The detailed description set forth below, together with the accompanying drawings, is intended to explain exemplary embodiments of the present invention and is not intended to represent the only embodiments in which the present invention may be practiced. In the drawings, portions irrelevant to the description may be omitted for clarity in describing the present invention, and the same reference numerals may be used throughout the specification for identical or similar components.

[0026] FIG. 1 is a schematic diagram illustrating the operation of an electronic device according to one embodiment of the present invention.

[0027] An electronic device (100) according to one embodiment of the present invention is a device that classifies whether an abdominal CT image (10) corresponds to one of normal, high-grade small bowel obstruction (HGSBO), and low-grade small bowel obstruction (LGSBO), and can be implemented as a computer, a programmable logic controller (PLC), a server, a smart phone, a tablet PC, a smart pad, a laptop, etc.

[0028] Hereinafter, high-risk small bowel obstruction is defined as small bowel obstruction requiring surgery, while low-risk small bowel obstruction is defined as small bowel obstruction amenable to non-surgical management. Furthermore, low-risk small bowel obstruction is considered to include paralytic ileus.

[0029] An abdominal CT image (10) is a 3D image obtained by performing computed tomography on the abdomen of a subject, and may include multiple segmented 2D images.

[0030] Classification information (20) is information that classifies whether the abdominal CT image (10) corresponds to normal, high-risk small bowel obstruction, or low-risk small bowel obstruction.

[0031] In the present invention, we propose a novel method for predicting high-risk small bowel obstruction using an optimized 3D-synthetic neural network.

[0032] More specifically, the present invention proposes a pooling and bifurcated structure classifier that maintains depth information of abdominal CT images to classify high-risk small bowel obstructions.

[0033] Hereinafter, the configuration and operation of an electronic device (100) according to one embodiment of the present invention will be specifically described with reference to the drawings.

[0034] FIG. 2 is a block diagram illustrating the configuration of an electronic device according to one embodiment of the present invention.

[0035] An electronic device (100) according to one embodiment of the present invention includes an input unit (110), a communication unit (120), a display unit (130), a storage unit (140), and a processor (150).

[0036] The input unit (110) generates input data in response to user input of the electronic device (100). For example, the user input may be a user input for driving the electronic device (100), and in addition, it may be applied without limitation to a user input required for obtaining classification information from an abdominal CT image.

[0037] The input unit (110) includes at least one input means. The input unit (110) may include a keyboard, a key pad, a dome switch, a touch panel, a touch key, a mouse, a menu button, etc.

[0038] The communication unit (120) can communicate with external devices such as CT equipment and medical systems to transmit and receive abdominal CT images, deep learning models, classification information, etc.

[0039] To this end, the communication unit (120) can perform wireless communication such as 5G (5th generation communication), LTE-A (long term evolution-advanced), LTE (long term evolution), Wi-Fi (wireless fidelity), Bluetooth, or wired communication such as LAN (local area network), WAN (Wide Area Network), and power line communication.

[0040] The display unit (130) displays display data according to the operation of the electronic device (100). The display unit (130) may display, for example, a screen displaying an abdominal CT image, a screen displaying classification information, a screen receiving user input, etc.

[0041] The display unit (130) includes a liquid crystal display (LCD), a light emitting diode (LED) display, an organic light emitting diode (OLED) display, a micro electro mechanical systems (MEMS) display, and an electronic paper display. The display unit (130) may be implemented as a touch screen by being combined with the input unit (110).

[0042] The storage unit (140) stores the operation programs of the electronic device (100). The storage unit (140) includes non-volatile storage that can preserve data (information) regardless of whether power is supplied, and volatile memory that cannot preserve data if power is not supplied and into which data to be processed by the processor (150) is loaded. The storage includes flash memory, hard-disc drive (HDD), solid-state drive (SSD), read-only memory (ROM), etc., and the memory includes buffer, random access memory (RAM), etc.

[0043] The storage unit (140) can store abdominal CT images, deep learning models, classification information, etc., and can store computational programs, etc., required in the process of obtaining classification information, extracting feature information, classifying abdominal CT images into either normal or abnormal, and classifying abdominal CT images into either high-risk small intestine obstruction or low-risk small intestine obstruction.

[0044] The processor (150) can control at least one other component (e.g., hardware or software component) of the electronic device (100) by executing software such as a program, and can perform various data processing or operations.

[0045] According to one embodiment of the present invention, a processor (150) may obtain classification information by inputting an abdominal CT image of a subject into a deep learning model trained to classify the abdominal CT image into one of normal, high-grade small bowel obstruction (HGSBO), and low-grade small bowel obstruction (LGSBO).

[0046] At this time, the processor (150) can learn a deep learning model or receive and store a deep learning model that has been learned and created from the outside and use it, but is not limited to either one.

[0047] Meanwhile, the processor (150) may perform at least a portion of the data analysis, processing, and result information generation for performing the above operations using at least one of a machine learning, neural network, or deep learning algorithm as a rule-based or artificial intelligence (AI) algorithm. Examples of the neural network may include models such as a CNN (Convolutional Neural Network), a DNN (Deep Neural Network), and an RNN (Recurrent Neural Network).

[0048] FIG. 3 is a diagram illustrating the structure of a deep learning model according to one embodiment of the present invention.

[0049] Referring to FIG. 3, the deep learning model (300) is composed of a feature extractor (310) that extracts feature information including CT depth information from an abdominal CT image (10), a classifier (320) (hereinafter referred to as a first classifier (320)) that classifies the abdominal CT image (10) into either normal or abnormal using the feature information, and a classifier (330) (hereinafter referred to as a second classifier (330)) that classifies the abdominal CT image into either high-risk small intestine obstruction or low-risk small intestine obstruction using the feature information in the case of abnormality.

[0050] As described above, the processor (150) can learn a deep learning model (300) or receive and store a deep learning model (300) generated through learning from an external source and use it. Hereinafter, the process by which the processor (150) learns a deep learning model (300) will be described.

[0051] First, abdominal CT images (10) corresponding to three groups, normal group, high-risk small bowel obstruction group, and low-risk small bowel obstruction group, are prepared. The abdominal CT image (10) is a 3D image with channel (C) Хhorizontal (W) Хvertical (H) Хdepth (D) dimensions.

[0052] The feature extractor (310) is a backbone network of the deep learning model (300), and is composed of a convolutional layer and a pooling layer. It can extract feature information of an abdominal CT image (10) by repeating convolution and pooling processes. The feature extractor (310) may be, for example, ResNet, ResNext, WideResNet, DenseNet, EfficientNet, etc., but is not limited to any one of them.

[0053] At this time, the processor (150) can extract feature information that maintains depth information of the abdominal CT image (10) through the feature extractor (310). The reason for maintaining depth information is as follows.

[0054] Most diseases are concentrated in extremely small areas of CT images. Therefore, retaining as much information as possible in the final pooling layer is essential to capture information from these small areas. However, most deep learning models compress the depth information of the input data in the pooling layer for various reasons, such as data volume and learning speed. For example, an input image of size CХHХWХD is pooled into CХ1Х1Х1 to compress the depth information.

[0055] That is, lesions such as high-risk small bowel obstructions and low-risk small bowel obstructions can only appear at specific depths in 3D abdominal CT images, making it crucial to learn depth-specific features. Therefore, the present invention proposes a method for retaining this depth information to gather information from small regions of 3D abdominal CT images. This is also known as Depth Retention Pooling (DRP).

[0056] The processor (150) can obtain feature information including CT depth information by performing pooling that preserves depth information for the vertical and horizontal spaces of the abdominal CT image through the feature extractor (310).

[0057] Feature information including CT depth information can be passed to the classifier through a fully connected layer.

[0058] Meanwhile, abdominal CT images of high-risk and low-risk small bowel obstructions are similar. Therefore, in order to more accurately distinguish between high-risk and low-risk small bowel obstructions, the present invention uses a dual-branch architecture (DBA).

[0059] First, the first classifier (320) is a basic classifier that can classify an abdominal CT image as either normal or abnormal.

[0060] The processor (150) can train the first classifier (320) to classify the abdominal CT image into either normal or abnormal using feature information of a normal abdominal CT image and an abnormal abdominal CT image. In this case, the abnormal abdominal CT image may include an abdominal CT image in the case of high-risk small bowel obstruction and an abdominal CT image in the case of low-risk small bowel obstruction.

[0061] At this time, the processor (150) can train the first classifier (320) to classify the abdominal CT image as either normal or abnormal by using compressed feature information of depth information of a normal abdominal CT image and an abnormal abdominal CT image.

[0062] In another embodiment, the processor (150) may train the first classifier (320) to classify the abdominal CT image as either normal or abnormal using feature information including CT depth information of a normal abdominal CT image and an abnormal abdominal CT image.

[0063] The second classifier (330) is a microclassifier capable of classifying abdominal CT images into either high-risk small bowel obstruction or low-risk small bowel obstruction. As previously described, the second classifier (330) utilizes depth information in abdominal CT images to distinguish between high-risk and low-risk small bowel obstruction.

[0064] The processor (150) can train the second classifier (330) to classify the abdominal CT image into either a high-risk small bowel obstruction or a low-risk small bowel obstruction using feature information including CT depth information of the abnormal abdominal CT image.

[0065] Classification information (20) is information output from the first classifier (320) and the second classifier (330), and may include probabilities for each group. However, the present invention is not limited thereto, and may include result values ​​based on probabilities.

[0066] If the first classifier (320) predicts that the input image is abnormal using the following mathematical expression 1, the probabilities of the two classifiers can be used.

[0067]

[0068] According to one embodiment of the present invention, a deep learning classifier can be used to effectively detect high-risk small bowel obstruction patients using only abdominal CT images.

[0069] A novel 3D classifier with a dual-branching structure and depth-preserving pooling based on the artificial intelligence algorithm of the present invention can be a reliable screening and decision support tool for high-risk patients with small bowel obstruction.

[0070] FIG. 4 is a diagram illustrating an operation flow diagram of an electronic device according to one embodiment of the present invention.

[0071] According to one embodiment of the present invention, the processor (150) can input an abdominal CT image of a subject into a deep learning model (S10).

[0072] The processor (150) can receive an abdominal CT image of a subject from a medical system such as CT equipment or an electronic health record (EHR), and the path for obtaining the abdominal CT image is not limited to any one.

[0073] The deep learning model is identical to the deep learning model described above with reference to Figure 3.

[0074] The processor (150) can extract feature information including CT depth information from an abdominal CT image of a subject through a feature extractor of a deep learning model (S20).

[0075] The processor (150) can classify the abdominal CT image of the subject into either normal or abnormal using feature information including CT depth information through the first classifier of the deep learning model (S30).

[0076] If the abdominal CT image of the subject is abnormal (No in S30), the processor (150) can classify the abdominal CT image into either a high-risk small bowel obstruction or a low-risk small bowel obstruction using feature information including CT depth information through the second classifier of the deep learning model (S40).

[0077] The processor (150) can obtain classification information according to the results classified through S30 and S40 (S50).

[0078] For example, if the abdominal CT image of the subject is normal (Yes in S30), the processor (150) can obtain classification information indicating that the subject is normal.

[0079] Alternatively, if the abdominal CT image of the subject is abnormal (No of S30) and the subject has a high-risk small bowel obstruction (Yes of S40), the processor (150) can obtain classification information indicating that the subject has a high-risk small bowel obstruction.

[0080] Alternatively, if the abdominal CT image of the subject is abnormal (No of S30) and has low-risk small bowel obstruction (No of S40), the processor (150) can obtain classification information indicating that the subject has low-risk small bowel obstruction.

[0081] Classification information can be expressed in various formats and is not limited to a single format. For example, for the first classifier, classification information may include probabilities for normal, high-risk small bowel obstruction, and low-risk small bowel obstruction, information regarding any one of the three probabilities exceeding a threshold, or information regarding the highest probability among the three. In this case, if either high-risk or low-risk small bowel obstruction is present, they can be grouped and output as abnormal.

[0082] For the second classifier, the classification information may include a probability for high-risk small bowel obstruction, low-risk small bowel obstruction, information about which of the two probabilities exceeds a threshold, or information about which of the two probabilities has a higher probability.

[0083] Elderly, sarcopenic or diabetic patients are more susceptible to progressive intestinal ischemia due to high-risk small bowel obstruction, which can lead to major morbidities such as acute kidney injury (AKI), heart damage, cardiovascular events or death. However, the present invention can prevent this by appropriately responding to high-risk small bowel obstruction patients at an early stage.

[0084] Additionally, early and appropriate surgery in patients with high-risk small bowel obstruction can avoid intestinal resection and prevent short bowel syndrome, which can occur when a large portion of the small intestine is removed.

[0085] FIG. 5 is a diagram illustrating a heat map using Grad-CAM of a deep learning model according to one embodiment of the present invention and an existing deep learning model.

[0086] Gradient-weighted class activation mapping (Grad-CAM) uses a heatmap to show the basis for a deep learning model's classification decisions.

[0087] Figure 5 is a heatmap using gradient-weighted class activation mapping for CT images of multiple patients, where the first row is normal, and the second and third rows are high-risk and low-risk small bowel obstructions, respectively.

[0088] A is a Grad-CAM of a pure deep learning model with WideResNet as a backbone network, and B is a Grad-CAM of a deep learning model applying the present invention with WideResNet as a backbone network.

[0089] Heatmaps of A show activation in similar regions across all classes (rows), indicating that the model is not accurately classifying each individual class.

[0090] In the normal class, the heatmap indicates activity primarily in the lower central abdomen. However, in the high-risk small bowel obstruction class, activity is concentrated in a specific region of the abdomen. In the low-risk small bowel obstruction class, the heatmap shows widespread activity across various regions of the abdomen. This demonstrates that the deep learning model utilizing our invention differentiates each class differently.

[0091] A 3D CNN deep learning model according to one embodiment of the present invention can detect high-risk small bowel obstruction patients with high accuracy and efficiency.

[0092] [Example]

[0093] The following dataset cleaning process was performed to identify high-risk groups who may require emergency surgery for bowel obstruction. Participants were divided into three groups: 1) non-surgical group (patients who visited the emergency department or were hospitalized for bowel obstruction between January 1, 2000, and December 31, 2021); 2) surgical group (patients who underwent adhesive dissection for ASBO during the same period); and 3) normal subjects (healthy individuals with no abnormal findings on abdominal CT during a health checkup in 2019).

[0094] At this time, to improve the classification performance for each group, the data set was selected based on selection criteria (patient age, medical staff opinion, etc.) or the number of data per group was balanced.

[0095] Selection criteria for inclusion in the dating set were as follows: age ≥18 years, CT image findings suggesting intestinal obstruction, and obstruction mechanism due to adhesions.

[0096] Prior to this procedure, we identified normal individuals (n=1,000) without abnormal abdominal CT findings from a health checkup conducted in 2019. Because the etiology and clinical aspects of intestinal obstruction are highly variable, a detailed review of the cases was necessary. Cases were excluded due to non-adhesive mechanisms, other anatomic locations, or pathologic conditions such as malignancy, peritonitis, or inflammatory bowel disease. Furthermore, cases with gastrografin use and postoperative intestinal obstruction within 1 month were excluded. Overt ischemia and necrosis of the small intestine were also excluded.

[0097] We identified patients with high-risk intestinal obstruction by examining the surgical and non-surgical groups. First, we used the ICD codes for intestinal obstruction in the non-surgical group. Cases where abdominal CT images were unavailable and cases of non-adhesive acute small bowel obstruction were excluded. High-risk small bowel obstruction was defined as complete or incomplete high-grade obstruction with a closed loop, adhesive band, or abrupt luminal stenosis. Low-risk small bowel obstruction was defined as fluid-filled distension of the small bowel with one of the following findings: low probability of obstruction, low-grade, or partial obstruction.

[0098] All patients in the surgical group who underwent abdominal surgery were evaluated at the hospital. Cases were subdivided based on criteria for adhesive delamination due to adhesive SBO. Finally, high-risk surgical findings, such as adhesive bands or closed loops, consistent with preoperative radiographic findings such as high-grade occlusion or suffocation, were included in the final cohort.

[0099] Ultimately, 209 high-risk small bowel obstructions, 119 low-risk small bowel obstructions, and 250 normal cases were obtained, and over 38,000 CT images were used. The dataset was divided into 462 training sets and 116 test sets to build a model, and five-fold cross-validation was performed.

[0100] During training, we used a stochastic gradient descent optimizer with a learning rate of 0.1, weight decay of 5e-4, and learning rate decay of 0.01, using a cosine annealing scheduler for 50 epochs. We also applied gradient clipping and efficient data augmentation methods such as flip, gamma, noise, motion, bias field, random affine, and elastic deformation. The input images were cropped and padded to 224х224х112 pixels after resampling the voxel space by 2.

[0101] Classification performance was evaluated based on accuracy, sensitivity, specificity, and area under the receiver operating characteristic curve.

[0102] Table 1 shows that the proposed network performed well in terms of accuracy, specificity, sensitivity, and AUROC. Furthermore, regardless of the backbone type, the deep learning model using the present invention always outperformed the baseline deep learning model.

[0103] After 5-fold cross-validation, the WideResNet classifier using a dual-branch architecture with depth-preserving pooling achieved an accuracy of 72.6% across all test sets, an area under the receiver-operating characteristic of 0.90, a sensitivity of 72.6%, a specificity of 86.3%, a positive predictive value of 74.1%, and a negative predictive value of 86.6%.

[0104] BackboneMethodAccuracySpecificitySensitivityAUROCResNetBasic66.1583.0866.150.848 ± 0.05This invention70.2685.1370.260.876 ± 0.02ResNextBasic68.4684.2368.460.874 ± 0.02This invention66.4183.2166.410.883 ± 0.01WideResNetBasic65.1382.5665.130.861 ± 0.02This invention72.5686.2872.560.896 ± 0.01DenseNetBasic75.1283.8362.810.868 ± 0.05This Invention 72.6883.2863.410.873 ± 0.07 EfcientNet Basic 71.2282.6460.410.841 ± 0.04 This invention 71.7183.1663.360.868 ± 0.03

Claims

1. In an electronic device for classifying small bowel obstruction using abdominal computed tomography (CT) images, A processor that inputs an abdominal CT image of a subject into a deep learning model trained to classify the abdominal CT image into one of normal, high-grade small bowel obstruction (HGSBO), and low-grade small bowel obstruction (LGSBO) to obtain classification information, The above deep learning model is, A feature extractor for extracting feature information including CT depth information from an abdominal CT image; A first classifier that classifies an abdominal CT image into either normal or abnormal using the above feature information; and An electronic device comprising a second classifier that classifies the abdominal CT image into either a high-risk small bowel obstruction or a low-risk small bowel obstruction using the above feature information, if abnormal.

2. In paragraph 1, The above processor, An electronic device that obtains feature information including CT depth information by performing pooling that preserves depth information for vertical and horizontal spaces of an abdominal CT image through the above feature extractor.

3. In paragraph 1, The above processor, An electronic device that trains a first classifier to classify an abdominal CT image as either normal or abnormal using feature information of a normal abdominal CT image and an abnormal abdominal CT image.

4. In paragraph 1, The above processor, An electronic device that trains a second classifier to classify an abdominal CT image into either a high-risk small bowel obstruction or a low-risk small bowel obstruction using feature information and CT depth information of an abnormal abdominal CT image.

5. A method for classifying small bowel obstruction using abdominal computed tomography (CT) images performed by an electronic device, A step of obtaining classification information by inputting an abdominal CT image of a subject into a deep learning model trained to classify the abdominal CT image into one of normal, high-grade small bowel obstruction (HGSBO), and low-grade small bowel obstruction (LGSBO), The above deep learning model is, A feature extractor for extracting feature information including CT depth information from an abdominal CT image; A first classifier that classifies an abdominal CT image into either normal or abnormal using the above feature information; and A method characterized in that, in case of abnormality, a second classifier is configured to classify the abdominal CT image into either a high-risk small bowel obstruction or a low-risk small bowel obstruction using the above characteristic information.

6. In paragraph 5, The step of obtaining the above classification information is: A method comprising a step of obtaining feature information including CT depth information by performing pooling that preserves depth information for vertical and horizontal spaces of an abdominal CT image through the feature extractor.

7. In paragraph 5, A method further comprising the step of training a first classifier to classify an abdominal CT image as either normal or abnormal using feature information of a normal abdominal CT image and an abnormal abdominal CT image.

8. In paragraph 5, A method further comprising the step of training a second classifier to classify the abdominal CT image into either a high-risk small bowel obstruction or a low-risk small bowel obstruction using feature information and CT depth information of the abnormal abdominal CT image.

Images