Multi-phase CT (Computed Tomography) image detection method for hepatic focal lesion

Abstract

The invention relates to the technical field of medical image processing, and discloses a multi-phase CT (Computed Tomography) image detection method for liver focal lesions, which comprises the following steps of: acquiring four-phase CT images, adjusting the size to unify the size, independently grouping the four-phase CT images, inputting four groups of RetinaNet multi-scale detection networks with shared parameters, extracting a feature map I grouped according to the phase, and extracting a feature map II grouped according to the phase; respectively inputting the four groups of feature maps I into an improved feature pyramid network to output feature maps II grouped according to phase, inputting the four groups of feature maps II into a ConvLSTM time sequence modeling module to carry out time sequence modeling, and finishing screening and outputting a result through a detection head and using non-maximum suppression. According to the method, through joint modeling of the multi-phase CT images, the dynamic evolution characteristics of the lesion in the time dimension are effectively extracted, and the recognition precision of the CT image of the liver lesion is remarkably improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of medical image processing, and particularly relates to a multi-phase CT image detection method for liver focal lesions. BACKGROUND

[0002] Liver failure is a common and critical liver dysfunction syndrome in clinical practice, which has complex causes and rapid progression, and is accompanied by pathological processes such as significant metabolic disorders, toxin accumulation and coagulation dysfunction. The clinical manifestations include jaundice, ascites, consciousness disorders and other multiple system symptoms. In the management of liver failure and related chronic liver diseases, early detection and accurate diagnosis of liver focal lesions (such as hepatocellular carcinoma, liver cysts, hemangiomas and focal nodular hyperplasia) are of great significance to improve the timeliness of intervention and the success rate of treatment.

[0003] Medical images, especially computed tomography (CT), play an irreplaceable role in liver lesion evaluation. Traditional methods often use single-phase CT images for lesion detection, but due to the significant enhancement mode differences of different types of liver lesions during contrast, single-phase images often fail to fully reflect the evolution characteristics of the lesions. For example, some malignant lesions show significant enhancement in the arterial phase and rapid washout in the portal venous phase, while benign lesions may still retain enhancement in the delayed phase. If only relying on images at a certain time point, it is easy to cause misjudgment or missed detection.

[0004] Current mainstream detection algorithms mostly use static modeling strategies, which only process a single image frame and cannot capture the density changes, morphological evolution and imaging feature changes of lesions between multiple phases. They lack the ability to model time dimension information. In addition, in actual clinical scanning, due to patient breathing, heartbeat or body position changes, there are often image offsets and organ deformations between multiple phases, resulting in significant registration errors in directly fused images. Traditional image processing algorithms are difficult to align these structural differences, which may introduce artifacts and even mask lesion features, limiting the effective use of multi-phase information. At the same time, with the popularization of high-resolution CT devices, the scanning layer thickness is decreasing and the number of images is increasing significantly, and radiologists are facing unprecedented work pressure. In the prior art, although there are methods for processing multi-phase CT images to improve the effect of target classification, the temporal relationship between phases is not considered, and there is a lot of room for improvement. SUMMARY

[0005] The technical problem to be solved by the present application is to provide a multi-phase CT image detection method for liver focal lesions, which can fuse multi-phase CT image sequences, has time sequence modeling and spatial alignment capabilities, and can classify and identify liver lesions in multi-phase CT images to improve detection accuracy.

[0006] To address the aforementioned technical problems, this invention provides a method for detecting focal liver lesions using multi-phase CT images, comprising the following steps: acquiring four-phase CT images, adjusting their size to a uniform value, inputting them into an offline-trained multi-phase fusion detection network, and outputting a detection image with binary classification results.

[0007] The multi-phase fusion detection network is an improvement based on the RetinaNet network, including a RetinaNet multi-scale detection network, an improved feature pyramid network, a ConvLSTM temporal modeling module, and a detection head. Four phase CT images are independently grouped and then input into four sets of RetinaNet multi-scale detection networks with shared parameters to extract the phase-grouped images. Feature maps, then four groups Each feature map is input to the improved feature pyramid network, and the output is grouped periodically. Feature maps, then four groups After the feature map is input into the ConvLSTM temporal modeling module for temporal modeling, the result is filtered and output after passing through the detection head and using non-maximum suppression.

[0008] An improvement to the multi-phase CT imaging method for detecting focal liver lesions according to the present invention:

[0009] The RetinaNet multi-scale detection network is an improvement based on the ResNet50 network. A self-attention module and a deformable convolution module are added to stage 3 of the original ResNet50 network, and a self-attention module is added to stages 4 and 5. The C3 feature map output from the original ResNet50 network is then passed through the self-attention module and the deformable convolution module sequentially before being output. Feature maps: The C4 and C5 feature maps output by the original ResNet50 network are each processed by a self-attention module to output... Feature maps and Feature map.

[0010] As a further improvement to the multi-phase CT imaging method for detecting focal liver lesions of the present invention:

[0011] The improved feature pyramid network is based on a feature pyramid network with added self-attention modules. The P3–P7 feature maps output by the feature pyramid network are each processed by the self-attention module before being output. Feature map.

[0012] As a further improvement to the multi-phase CT imaging method for detecting focal liver lesions of the present invention:

[0013] The ConvLSTM time series modeling module includes a ConvLSTM network. The feature maps are arranged into a temporal feature sequence, which is then passed through four ConvLSTM networks with shared weight parameters, and then concatenated and grouped convolutions are performed in the time dimension.

[0014] As a further improvement to the multi-phase CT imaging method for detecting focal liver lesions of the present invention:

[0015] The offline training process of the multi-phase fusion detection network is as follows:

[0016] Four-phase CT images of the liver were acquired, including the plain scan phase, arterial phase, portal venous phase, and delayed phase. After manually labeling the regions and tags of focal liver lesions, the image sizes were uniformly adjusted. Random flipping, region cropping, and Gaussian noise were applied, and the images were randomly divided into training and testing sets at a 4:1 ratio. The training set was input into the multi-phase fusion detection network, and the training process included forward propagation, backward propagation, and weight updates. The performance was evaluated using five-fold cross-validation.

[0017] The beneficial effects of this invention are mainly reflected in:

[0018] 1. This invention effectively extracts the dynamic evolution features of lesions in the time dimension by jointly modeling multi-phase CT images, significantly improving the recognition accuracy of CT images of liver lesions;

[0019] 2. This invention preserves the multi-phase sequence dependence by introducing a ConvLSTM structure, which helps the model model temporal context features and improves the detection performance and accuracy of multi-phase CT images of liver lesions; the self-attention mechanism enhances the modeling of spatial saliency between phases, guides the model to focus on key areas, and effectively suppresses background interference; the deformable convolution module dynamically aligns the phase offset regions, improving the model's robustness to lesion morphological variations and unregistered images;

[0020] 3. The multi-phase fusion detection network structure of the present invention has good detection performance, with an AP@0.5 index of 0.7982, and has strong clinical online use and auxiliary diagnostic application value. Attached Figure Description

[0021] The specific embodiments of the present invention will be further described in detail below with reference to the accompanying drawings.

[0022] Figure 1 This is a schematic diagram of the RetinaNet basic network structure;

[0023] Figure 2 This is a schematic diagram of the ConvLSTM timing modeling module in the multi-phase fusion detection network of the present invention;

[0024] Figure 3 This is a schematic diagram of a self-attention module;

[0025] Figure 4 This is a schematic diagram of a variable convolutional module;

[0026] Figure 5 This is a schematic diagram of the multi-phase fusion detection network of the present invention;

[0027] Figure 6 This is an example image showing the online detection results of a multi-phase CT imaging method for detecting focal liver lesions according to the present invention. Detailed Implementation

[0028] The present invention will be further described below with reference to specific embodiments, but the scope of protection of the present invention is not limited thereto:

[0029] Example 1: A method for detecting focal liver lesions using multi-phase CT imaging, such as... Figures 1-6 As shown, the specific process is as follows:

[0030] 1. Acquisition and preprocessing of multi-phase CT image data

[0031] Liver CT images from 724 subjects obtained from collaborating hospitals were collected. Each liver CT image included four phases: plain (NC), arterial (ART), portal venous (PV), and delayed (DL), totaling 2896 images. All images underwent anonymization, and experts annotated the areas of focal liver lesions, manually adding labels including the presence of focal liver lesions and bounding box coordinates. All phase images were uniformly resized to 224×224 pixels and enhancement strategies such as random flipping, region cropping, and Gaussian noise were applied.

[0032] 2. Construct a multi-phase fusion detection network

[0033] This invention uses the RetinaNet network as the basic detection framework to construct a multi-phase fusion detection network for the detection of multi-phase CT images. It includes a RetinaNet multi-scale detection network, an improved feature pyramid network, a ConvLSTM temporal modeling module, and a detection head network. Images of each phase (temporally, plain scan, arterial phase, portal venous phase, and delayed phase) are independently grouped, and the RetinaNet multi-scale detection network, with four sets of shared parameters, is used to extract the images. Feature maps, then output by an improved feature pyramid network. Feature maps, then the four groups of independently extracted features The feature map is input into the ConvLSTM temporal modeling module for temporal modeling. The final detection head outputs the category and bounding box of multi-phase CT images of liver lesions, and non-maximum suppression is used for filtering. This network achieves high-precision detection of multi-phase CT images through phase sequence modeling, multi-scale fusion, and structural optimization.

[0034] 2.1 Constructing the RetinaNet basic network

[0035] The RetinaNet basic network mainly consists of a feature extraction network (using the ResNet50 network), a feature fusion network (i.e., the feature pyramid network, FPN), and a detection head network (classification and regression detection head).

[0036] The ResNet50 network consists of stages 1-5. The input image is processed through stages 3-5 of ResNet50 to extract C3–C5 feature maps. These are then upsampled from top to bottom and fused laterally by a Feature Fusion Network (FPN) to generate five feature maps at scales P3–P7. These P3 and P4 feature maps are then input into the detection head network for classification and bounding box prediction. The P3 feature map is 28×28×256 in size, the P4 feature map is 14×14×256, the P5 feature map is 7×7×256, the P6 feature map is 4×4×256, and the P7 feature map is 2×2×256.

[0037] The input image size is 224×224. Multi-scale P3–P7 feature maps are used for object classification and bounding box prediction. Each anchor point has 9 preset anchor boxes with varying scales. Aspect Ratio FocalLoss was used to address the class imbalance problem, and Smooth L1 regression loss was adopted.

[0038] 2.2 Multi-phase fusion detection network

[0039] Based on the RetinaNet basic network described in step 2.1, this invention systematically improves the feature extraction, feature fusion, and temporal fusion aspects of multi-phase CT images. This includes improving the original feature extraction network (ResNet50) into a RetinaNet multi-scale detection network using a self-attention module (SA) and a deformable convolution module (DConv), improving the feature pyramid using a self-attention module (SA), and adding a ConvLSTM temporal modeling module before the classification and regression detection heads. This allows the network to maintain the efficiency of single-stage detection while fully exploring the spatiotemporal complementary information between phases.

[0040] 2.2.1 RetinaNet Multiscale Detection Network

[0041] Based on the existing feature extraction network (ResNet50 network), this invention enhances the C3, C4, and C5 features output by the ResNet50 network before feeding them into the improved feature pyramid network of this invention. Specifically, a self-attention module (SA) and a deformable convolutional module (DConv) are added to stage 3 of the original ResNet50 network. The original C3 feature map is output sequentially through the self-attention module (SA) and the deformable convolutional module (DConv). Feature maps are used to independently predict migrations for different phase channel groups. The DConv module is collaboratively supervised by the self-attention gating map and self-attention feature map output from the SA module for migration learning. The self-attention gating map guides the DConv module to focus on significant changes in the lesion region, while the self-attention feature map weights the regions of interest, further enhancing feature representation capabilities. Self-attention modules are added to stages 4 and 5 of the original ResNet50 network. The C4 and C5 feature maps output from the original stages 4 and 5 are processed by the SA module. Feature maps and The feature maps are then fed into an improved feature pyramid network. This improvement can alleviate the problem of interphase misalignment caused by changes in respiration or body position, and enhance the model's robustness to fine-grained geometric deformations.

[0042] Self-attention module such as Figure 3 As shown, the intermediate layer feature maps of the four phases are concatenated and then input into the self-attention module. This module generates queries through 1×1 convolutions. ,key ,value Then calculate the attention weights. Weighted generation of feature maps The final output is ,in For residual gating factor, To output weights, For the input feature map, This module produces the output feature map. It enhances the spatial saliency representation and fuses it with the original features through a trainable residual gating mechanism. Embedded in the middle of RetinaNet and the front of ConvLSTM, this module is used for saliency enhancement and cross-phase context modeling.

[0043] Variable convolutional modules such as Figure 4 As shown, to address the potential spatial offset between images from different time periods, this invention introduces a variable convolution module. The standard convolution location set is... Deformable convolution introduces offset Sampling location is ,in This is the current center position of the convolution. This represents the standard convolution sampling location. The offset is predicted via convolution, with input including the feature map and the output of the self-attention module; a 3×3 convolution kernel is used, and the output channels are consistent with the original features, enhancing the alignment effect between phases. This module dynamically predicts the spatial offset at each location. The standard convolution sampling positions are adjusted. Guided by the output of the self-attention module, the offset direction becomes more context-aware. This module is inserted into the middle layer (after C3) of the RetinaNet feature extraction network for lesion spatial alignment.

[0044] The self-attention module (SA) enhances feature representation, and the variable convolutional module (DConv) performs local structure alignment.

[0045] 2.2.2 Improvement of the Feature Pyramid

[0046] Based on the feature pyramid network of the basic RetinaNet network, self-attention modules (SA) are superimposed on the multi-scale feature maps of P3–P7 to obtain... Feature map. This improvement enhances the significant regions of the transphase phase and makes small lesion features more prominent.

[0047] 2.2.3 Add ConvLSTM timing modeling module

[0048] ConvLSTM timing modeling module, such as Figure 2 As shown, the improved feature pyramid network outputs each phase at each scale. The feature maps are arranged into a temporal feature sequence and input into a ConvLSTM network with four shared weight parameters. The outputs of all time steps are retained and concatenated in the time dimension. After 1×1 group convolution, the temporal fusion feature map is output, which extracts the temporal information of multi-phase images.

[0049] 2.3 Overall Structure of Multi-Phase Phase Fusion Detection Network

[0050] like Figure 5 As shown, the complete network includes: a RetinaNet multi-scale detection network, an improved feature pyramid network, a ConvLSTM temporal modeling module, and a detection head. The complete detection process involves the RetinaNet multi-scale detection network, where each phase image is independently input into four sets of shared parameters. The input size for each of the four phases is 224×224×3. Each phase is independently processed by convolution through a residual structure, sharing parameters but with separate channel groups to avoid interference between features from different phases. The phase-grouped images obtained after processing by the RetinaNet multi-scale detection network... The feature maps have sizes of 28×28×512, 14×14×1024, and 7×7×2048, respectively. Then, four groups... Each feature map is input to the improved FPN module, which outputs four groups of periodically grouped features. Feature map The feature map sizes are 28×28×256, 14×14×256, 7×7×256, 4×4×256, and 2×2×256, respectively. Then, four groups... The feature maps are sequentially input into the ConvLSTM temporal modeling module for temporal modeling and output a temporal fusion feature map. Finally, the detection head outputs the category and bounding box of multi-phase liver lesion CT images, and non-maximum suppression is used for filtering. This network achieves high-precision detection of multi-phase liver lesion CT images through phase sequence modeling, multi-scale fusion, and structural optimization.

[0051] 3. Network Training and Testing

[0052] The preprocessed phase images from step 1 were divided into training and testing sets in a 4:1 ratio, with four multi-phase images from the same patient constituting one sample. The AdamW optimizer was used with an initial learning rate of 0.001, a batch size of 4, and an epoch count of 200. Cosine annealing and warm-up strategies were employed to dynamically adjust the learning rate. The training process included forward propagation, backpropagation, and weight updates. After training, the model parameters were saved as a deployable model file. During the testing phase, the multi-phase images were input into the trained network, which output binary classification detection results for multi-phase CT images of liver lesions, including class probabilities and bounding box coordinates. After non-maximum suppression, the results were visualized and their accuracy evaluated. Five-fold cross-validation was used in the model testing phase, with the final metric being the AP@0.5 mean of the five-fold cross-validation.

[0053] 4. Online use

[0054] The four-phase CT images of new subjects were resized (uniformed to 224×224) and input into a pre-trained multi-phase fusion detection network. After multi-scale feature extraction, temporal modeling, attention enhancement, and spatial alignment, the detection head outputs binary classification results for multi-phase CT image recognition, including confidence scores and bounding box coordinates. Finally, NMS filtering is used to generate visualized detection results, such as... Figure 6 As shown, this online process requires only one forward inference to complete the entire detection process, enabling rapid, automated, and repeatable identification and detection of multi-phase CT images of localized liver lesions.

[0055] 5. Experiment

[0056] The experimental environment is shown in Table 1 below.

[0057] Table 1: Experimental Environment Configuration Parameters

[0058] Environment configuration Parameter Operating system CentOS 7.3.1611 Processor Intel Xeon 12 * E5-2609 v3 @ 1.9GHz, 15M Cache Memory 256 Graphics card NVIDIA Tesla P4 16GB Development environment Python 3.8 + PyTorch 1.13

[0059] 5.1 Analysis and Experiment of Multi-Phase Phase Fusion Detection Algorithm

[0060] To verify the effectiveness of multi-phase CT image input, this experiment designed multiple phase combination strategies and input them into a detection network capable of processing multi-phase data and adapting to single-phase input for comparative analysis of detection performance. In this experiment, AP@0.5 represents the average accuracy (AP) when the Intersection over Union (IoU) threshold is 0.5. The experimental data source is the liver CT images of 724 patients acquired in step 1 of Example 1, categorized into four phases: plain phase (NC), arterial phase (ART), portal venous phase (PV), and delayed phase (DL). The experimental results are shown in Table 2.

[0061] Table 2: Experimental Results of Multi-Phase Phase Fusion Detection Algorithm

[0062] Input data Precision Recall AP@0.5 NC 0.6716 0.6522 0.6618 ART 0.7424 0.7101 0.7283 PV 0.7576 0.7246 0.7553 DL 0.6970 0.6667 0.6815 NC+ART 0.7538 0.7101 0.7414 ART+PV 0.7846 0.7391 0.7758 NC+ART+PV 0.7879 0.7536 0.7840 4-phase complete input 0.8030 0.7681 0.7982

[0063] The networks used for inputting single-phase CT images in the first to fourth rows of Table 2 include a serially connected RetinaNet multi-scale detection network, an improved feature pyramid network, and a ConvLSTM network (which at this point only serves as a fusion network). (The role of multi-scale feature maps) and the detection head. The network inputting the biphasic CT images in rows 5 and 6 includes two sets of RetinaNet multi-scale detection networks with shared parameters, extracting features classified by phase. Feature maps are then used, and the improved feature pyramid network outputs phase-based classifications. Feature maps, then two groups The feature maps are input into two ConvLSTM networks with shared weight parameters, then concatenated and grouped convolutions are performed along the time dimension, and finally the detection head outputs binary classification results. Similarly, the network inputting the seven rows of three-phase CT images includes three sets of RetinaNet multi-scale detection networks with shared parameters, which extract classifications according to phase. Feature maps are then used, and the improved feature pyramid network outputs phase-based classifications. Feature maps, then 3 groups The feature maps are input into three ConvLSTM networks with shared weight parameters, then concatenated and grouped convolutions are performed along the time dimension. Finally, the detection head outputs the binary classification result. The network with complete input of four phases in the eighth row is the multi-phase fusion detection network of this invention.

[0064] As shown in Table 2, among the four single-phase inputs, the PV phase image performed best, followed by the ART phase. This aligns with clinical observations that most liver lesions exhibit the clearest boundaries and highest background tissue contrast in the PV phase, which is beneficial for lesion detection. The NC and DL phases, however, performed significantly worse than the significant phases, indicating that the model struggles to capture significant features and performs relatively poorly in stages where lesions have not significantly enhanced or have already shown enhancement decline.

[0065] In the phase combination experiment, the combined input of multiple phases showed significant complementarity, effectively improving detection accuracy. Specifically, the AP@0.5 of ART + PV reached 77.58%, an improvement of 3.75% compared to ART and 2.05% compared to PV, indicating that the combined input of multiple phase images can capture the manifestation of lesions at different enhancement points, providing richer semantic and structural information. Further combination of NC + ART + PV increased AP@0.5 to 78.40%, indicating that plain scan images still provide some supplementary information.

[0066] The multi-phase fusion detection network of this invention, based on input NC, ART, PV, and DL phase images, introduces ConvLSTM to model temporal features, a self-attention module to achieve cross-phase interaction, and a deformable convolution module to enhance geometric alignment capabilities, achieving optimal overall detection performance. AP@0.5 reaches 79.82%, with precision and recall of 80.30% and 76.81%, respectively. Experiments verify the significant advantages of the multi-phase fusion strategy in the detection of multi-phase CT images of focal liver lesions, and also demonstrate that the multi-scale fusion and temporal modeling network proposed in this patent possesses good detection enhancement capabilities and has potential online application prospects.

[0067] 5.2 Ablation Test

[0068] This invention is based on the RetinaNet ResNet50 network. It optimizes the detection network by introducing a self-attention module and a deformable convolution module to construct a RetinaNet multi-scale detection network. A ConvLSTM temporal modeling module is built using a ConvLSTM network, and an improved feature pyramid network is constructed by adding a self-attention module to the feature pyramid network. This enhances the global modeling capability and local geometric adaptability of feature extraction. To verify the effectiveness of the improved network, an ablation experiment was designed for comparative analysis. The experimental data source was uniformly the liver CT images of 724 patients acquired in step 1 of Example 1. The experimental results are shown in Table 3.

[0069] Table 3: Ablation Experiment Results

[0070] Model Precision Recall AP@0.5 RetinaNet 0.7656 0.7101 0.7424 RetinaNet + ConvLSTM time series modeling module 0.7813 0.7246 0.7665 RetinaNet + ConvLSTM time series modeling module + SA 0.7910 0.7681 0.7875 RetinaNet + ConvLSTM time series modeling module + DConv 0.7969 0.7391 0.7838 The present invention 0.8030 0.7681 0.7982

[0071] The experimental results show that the benchmark model RetinaNet, as the comparison baseline, has an AP@0.5 of 74.24%. Although it has good multi-scale modeling capabilities, it is insufficient in modeling dynamic features in multi-phase images and has difficulty capturing the evolution information of lesions between different phases.

[0072] The second row of Table 3 shows that when the ConvLSTM time-series modeling module was added to the baseline model, the model AP@0.5 increased to 76.65%, indicating that ConvLSTM can effectively utilize the time-series dependencies in the phase sequence, enhance the model's ability to capture the dynamic manifestations of lesions, thereby reducing the false negative rate and improving the detection coverage.

[0073] The third row of Table 3 further introduces a Self Attention (SA) module based on the network in the second row of Table 3. Specifically, it describes the outputs of the C3-C5 feature maps from the original ResNet50 network's stage 3-stage 5 after processing by the SA module. 3- 5. Feature maps: SA modules are superimposed on the P3–P7 multi-scale feature maps output by the feature pyramid network of the base RetinaNet network to obtain the feature maps. Feature maps. The introduction of the self-attention module improved AP@0.5 to 78.75% and further increased recall to 76.81%. This module effectively enhanced the information interaction and saliency modeling between multi-phase features by capturing long-range dependencies and salient regions between phases, making the model focus more on key regions and thus improving the overall detection capability.

[0074] The fourth row of Table 3 adds a deformable convolutional module (DConv) to the network independently based on the second row of Table 3. Specifically, a deformable convolutional module (DConv) jointly supervised by a self-attention gating map and a self-attention feature map is added to stage 3 of the original ResNet50 network. Although the introduction of this module increases the AP@0.5 to 78.38%, the recall rate is not significantly improved. This result indicates that DConv can enhance the model's adaptability to geometric offsets and unregistered regions, but its offset learning may be unstable in the absence of global guidance, and its contribution to recall is limited.

[0075] Ultimately, the multi-phase fusion detection network of this invention achieved the highest precision and recall, at 80.30% and 76.81%, respectively, with AP@0.5 increasing to 79.82%. This demonstrates that the synergistic effect of the two technologies can achieve unified optimization of global saliency modeling and local deformation alignment, significantly improving the model's robustness in complex phase images. Experimental results show that the improvements proposed in this patent bring considerable performance enhancements, verifying the effectiveness of temporal modeling and spatial structure optimization in improving the detection accuracy of multi-phase CT images of focal liver lesions.

[0076] (3) Experiment 3: Literature Comparison Experiment

[0077] To further verify the effectiveness of the multi-phase fusion detection network proposed in this patent, advanced algorithms in the field of medical image target detection were selected for reproduction. The liver CT image dataset of 724 patients collected in step 1 of Example 1 was used for performance comparison analysis. Each patient included four phases: plain scan phase (NC), arterial phase (ART), portal venous phase (PV), and delayed phase (DL). The input of the models of the comparison algorithms in various literatures can support the processing of four phases.

[0078] The experimental results are shown in Table 4.

[0079] Table 4: Comparison of experimental results with literature

[0080] Model Precision Recall AP@0.5 Document 1 0.6923 0.6522 0.6717 Document 2 0.7272 0.6957 0.7111 Document 3 0.7462 0.7246 0.7398 Document 4 0.7612 0.7391 0.7572 The model of the present invention 0.8030 0.7681 0.7982

[0081] The experimental results show that the traditional detection algorithm SSD used in Reference 1 performed the weakest. Its structure is based on static multi-scale features, which cannot effectively extract dynamic semantic information between different phases. Furthermore, its ability to model lesion boundaries is limited, resulting in a recall rate of only 65.22% and a high false negative rate.

[0082] Reference 2 uses Faster R-CNN, a classic two-stage detection framework. Although it is more accurate in generating candidate regions, with AP@0.5 increasing to 71.11%, it lacks the ability to explicitly model the phase information of the input image, making it difficult to fully mine time series features.

[0083] Reference 3 employs Grouped SSD (GSSD), introducing a multi-flow structure for feature extraction and fusion of multi-phase CT images. Its performance is significantly better than the two methods mentioned above, achieving an AP@0.5 of 73.98%. This improvement is attributed to GSSD's independent processing of features for each phase and its post-fusion mechanism. However, its direct use of 1×1 convolution for fusion, while alleviating the detection difficulties caused by phase mismatches to some extent, does not consider the temporal relationships between phases, thus limiting its potential for further improvement.

[0084] Reference 4 employs PixelLink, which, although initially designed for text detection tasks, possesses strong localization capabilities and the ability to handle overlapping targets due to its instance segmentation approach and boundary association mechanism. In this experiment, the method also demonstrates good robustness in medical image scenarios, achieving an AP@0.5 of 75.72%.

[0085] In contrast, the model proposed in this patent integrates four strategies: multi-scale detection, temporal modeling, spatial interaction, and geometric alignment. While fully modeling multi-phase information, it effectively solves problems such as phase misalignment and boundary ambiguity, thus achieving superior detection accuracy. Furthermore, the differences in metrics among the models also demonstrate that the simultaneous improvement in precision and recall indicates that the model reduces false positives while enhancing the overall ability to identify lesion regions in CT images. The above comparative experimental analysis proves that the multi-phase fusion detection network proposed in this patent is effective and advanced, providing more objective and efficient technical support for the auxiliary diagnosis and treatment of liver failure patients, and has strong clinical application value.

[0086] See reference 1: Liu W, Anguelov D, Erhan D, et al. Ssd: Single shot multibox detector[C] / / Computer Vision–ECCV 2016: 14th European Conference, Amsterdam, TheNetherlands, October 11–14, 2016, Proceedings, Part I 14. 2016: 21-37.

[0087] See reference 2: Ren S, He K, Girshick R, et al. Faster r-cnn: Towards real-timeobject detection with region proposal networks[J]. Advances in neuralinformation processing systems, 2015, 28.

[0088] See reference 3: Lee SG, Kim E, Bae JS, et al. Robust end-to-end focal liver lesion detection using un-registered multiphase computed tomography images[J]. IEEETransactions on Emerging Topics in Computational Intelligence, 2021, 7(2):319-329.

[0089] See reference 4: Deng D, Liu H, Li X, et al. Pixellink: Detecting scene text viainstance segmentation[C] / / Proceedings of the AAAI conference on artificialintelligence: vol. 32: 1. 2018.

[0090] Finally, it should be noted that the above examples are merely some specific embodiments of the present invention. Obviously, the present invention is not limited to the above embodiments and many variations are possible. All variations that can be directly derived or conceived by those skilled in the art from the disclosure of the present invention should be considered within the scope of protection of the present invention.

Claims

1. A method for detecting focal liver lesions using multi-phase CT imaging, characterized in that: The specific methods include the following: acquiring four phase CT images, adjusting and unifying their size, inputting them into an offline-trained multi-phase fusion detection network, and outputting detection images with binary classification results; The multi-phase fusion detection network is an improvement based on the RetinaNet network, including a RetinaNet multi-scale detection network, an improved feature pyramid network, a ConvLSTM temporal modeling module, and a detection head. Four phase CT images are independently grouped and then input into four sets of RetinaNet multi-scale detection networks with shared parameters to extract the phase-grouped data. Feature maps, then four groups Each feature map is input to the improved feature pyramid network, and the output is grouped periodically. Feature maps, then four groups After the feature map is input into the ConvLSTM temporal modeling module for temporal modeling, the result is filtered and output after passing through the detection head and using non-maximum suppression.

2. The method for detecting focal liver lesions using multi-phase CT imaging according to claim 1, characterized in that: The RetinaNet multi-scale detection network is an improvement based on the ResNet50 network. A self-attention module and a deformable convolution module are added to stage 3 of the original ResNet50 network, and a self-attention module is added to stages 4 and 5. The C3 feature map output from the original ResNet50 network is then passed through the self-attention module and the deformable convolution module sequentially before being output. Feature maps: The C4 and C5 feature maps output by the original ResNet50 network are each processed by a self-attention module to output... Feature maps and Feature map.

3. The method for detecting focal liver lesions using multi-phase CT imaging according to claim 2, characterized in that: The improved feature pyramid network is based on a feature pyramid network with added self-attention modules. The P3–P7 feature maps output by the feature pyramid network are each processed by the self-attention module before being output. Feature map.

4. The method for detecting focal liver lesions using multi-phase CT imaging according to claim 3, characterized in that: The ConvLSTM time series modeling module includes a ConvLSTM network. The feature maps are arranged into a temporal feature sequence, which is then passed through four ConvLSTM networks with shared weight parameters, and then concatenated and grouped convolutions are performed in the time dimension.

5. The method for detecting focal liver lesions using multi-phase CT imaging according to claim 4, characterized in that: The offline training process of the multi-phase fusion detection network is as follows: Four-phase CT images of the liver were acquired, including the plain scan phase, arterial phase, portal venous phase, and delayed phase. After manually labeling the regions and tags of focal lesions in the liver, the image size was uniformly adjusted. Random flipping, region cropping, and Gaussian noise were applied, and the images were randomly divided into training and test sets in a 4:1 ratio. The training set was input into the multi-phase fusion detection network, and the training process included forward propagation, backward propagation, and weight updates. The performance was evaluated using five-fold cross-validation during the testing process.

Images