Automatic segmentation method for brain tumours in computed tomography

Abstract

The invention relates to an automatic segmentation method for brain tumours in computed tomography, which aims to solve a number of important problems by ensuring accurate and effective segmentation of glioma tumours.

Description

[0001] AUTOMATIC SEGMENTATION METHOD FOR BRAIN TUMOURS IN COMPUTED TOMOGRAPHY

[0002] Technical Field of the Invention

[0003] The invention relates to an automatic segmentation method for brain tumours in computed tomography, which aims to solve a number of important problems by ensuring accurate and effective segmentation of glioma tumours.

[0004] State of the Art

[0005] Computed tomography (CT) provides a significant incentive in the segmentation of brain tumours. CT scans play a critical role in determining the location, size, and shape of tumours by presenting cross-sectional images of brain tissues. This helps doctors better understand the spread of tumours and the affected regions. However, segmentation of brain tumours is often a challenging process, since tumours are usually mixed with surrounding tissues and exhibit heterogeneous structures. This complexity gives rise to various difficulties in both manual and automatic segmentation methods.

[0006] Manual segmentation is generally a process carried out by experienced radiologists, and visual inspection and drawing tools are used to determine the boundaries of tumours. However, this method is time-consuming and prone to human error. Furthermore, the applicability of such manual processes is limited in large data sets. Therefore, automatic segmentation methods have emerged as an important alternative, especially in cases where large-scale data sets need to be processed rapidly and consistently.

[0007] Automatic segmentation techniques have achieved significant progress with the development of machine learning and deep learning algorithms. Convolutional neural networks (CNNs) come to the forefront in this context, since these algorithms use a large training data set to effectively classify and segment tumours in CT images. CNNs provide high accuracy rates by using multi-layered structures to determine the size, shape, and boundaries of tumours. These methods can improve their accuracy over time as the model is trained with more data.

[0008] In addition to deep learning techniques, various preprocessing and post-processing techniques are also applied to increase the accuracy of segmentation results. The preprocessing steps include denoising the images, enhancing the contrast, and normalisation. These steps improve the performance of segmentation algorithms by ensuring that tumours are clearly separated. The post-processing techniques aim to correct segmentation results, refine the boundaries, and increase accuracy. These processes ensure that the results become more reliable for clinical applications.

[0009] Another important development concerns the possibilities related to the integration of multimodal imaging data. A single CT scan may not usually reflect the full characteristics of tumours; therefore, combining it with additional imaging techniques such as magnetic resonance imaging (MRI) provides a more comprehensive assessment. Multimodal data allow tumours to be defined and segmented more accurately, thereby emphasising the need for a multidisciplinary approach.

[0010] Consequently, CT-based brain tumour segmentation is continuously evolving through both manual and automatic methods. Deep learning and other advanced techniques increase the accuracy and efficiency of segmentation, while preprocessing and postprocessing steps make these processes more reliable. These advances make a great contribution to the management and treatment of brain tumours in clinical applications, allowing patients to achieve better outcomes and treatment processes to be conducted more effectively.

[0011] Although various proposals and applications have been developed in the prior art for the segmentation of brain tumours in computed tomography, these developments remain insufficient. Some patent applications developed for this purpose are provided below.

[0012] In the prior art, patent application number CN111192245A mentions a network and method developed for brain tumour segmentation. The invention uses DCU-Net, a U- Net-based segmentation network. In the structure of the network, there is a spatial pyramid pooling structure added to the end of the contraction path, as well as dilated (hole) convolutions at various scales. The dilated convolutions are integrated into the skip connections of the network, which helps the network to understand a wider area. In addition, residual blocks created using the dilated convolutions and the original input have increased the efficiency of the network and improved the learning of features. The feature maps are expanded in the contraction path and combined with the corresponding stage of the expansion path. The segmentation method comprises steps such as cropping and preprocessing of the training data set, constructing the DCU-Net-based segmentation network, inputting the processed two-dimensional images into the model, obtaining the optimal parameters, and performing segmentation of the tumour region by inputting the images in the test data set into the segmentation model. This method aims to effectively solve problems such as over-segmentation and under-segmentation in brain tumour segmentation and to improve segmentation accuracy. The mentioned invention is designed to increase accuracy in brain tumour segmentation and to address problems related to segmentation.

[0013] In the prior art, patent application number CN109754404A covers an end-to-end tumour segmentation method developed based on a multi-attention mechanism. The method mainly consists of a backbone network part and an attention module part. The backbone network includes three sub-networks, and these sub-networks consist of improved 3D Residual ll-Net structures. The attention module has a specially designed double-branch structure. This method can overcome deficiencies of previous techniques, such as low training efficiency and low segmentation accuracy. In addition, it converts the problem of multiple tumour subregion segmentation into a series of binary segmentation tasks. The attention mechanism uses the segmentation result of the oedema region around the tumour as soft attention, and this soft attention is added to the subtask of segmenting the core part of the tumour. Moreover, the segmentation result of the core part of the tumour is incorporated, through the attention mechanism, into the segmentation subtask for the enhancement region within the tumour core. This method is suitable for the segmentation of 3D tumour lesion tissues with similar hierarchical structures, such as MRI and CT images, including brain tumours, and provides more accurate segmentation results. The mentioned invention is designed to improve accuracy in tumour segmentation and to enhance training efficiency.

[0014] In the prior art, patent application number TR2023017024A2 covers a system that enables brain tumours to be identified rapidly and accurately by means of artificial intelligence. The system analyses the data obtained from the MRI (Magnetic Resonance Imaging) or CT (Computed Tomography) scans of patients and determines the type, size, and location of brain tumours. By using deep learning-based methods, classification and segmentation of tumour images are achieved; in addition, the shape and volume of the tumour are measured through 3D modelling. The system comprises an electronic device and a server. The electronic device exchanges data using remote communication protocols and runs image processing applications. The server processes the images by means of medical imaging software and deep learning algorithms, performs detailed analysis of the tumours, and presents the results to healthcare professionals. This invention increases accuracy in the diagnosis of brain tumours, enabling patients to be treated more effectively and providing important information for treatment planning.

[0015] In the prior art, there is a need for an automatic segmentation method for brain tumours in computed tomography that comprises automatic segmentation algorithms, data preprocessing, training and validation, precise and rapid segmentation, and scalability features.

[0016] As a result, due to the drawbacks mentioned above and the inadequacy of current solutions regarding the subject matter, a development in the relevant technical field has become necessary.

[0017] The Aim of the Invention

[0018] The main aim of the invention is to ensure accurate and effective segmentation of glioma tumours.

[0019] Another aim of the invention is to provide an objective evaluation by means of computer-assisted segmentation. Thus, it reduces inconsistencies arising from the subjective interpretations of the expert and ensures that the tumour boundaries are determined more accurately and consistently.

[0020] Another aim of the invention is to accelerate the process significantly by means of automatic segmentation algorithms. Thus, it ensures that the tumour regions are determined quickly and efficiently. Another aim of the invention is to enable the analysis of more patient data or image sets. Thus, computer-assisted automatic segmentation ensures a significant acceleration of the segmentation process and increases efficiency.

[0021] Another aim of the invention is to help preserve important details during information propagation through mutual connections between the encoder and decoder by means of the developed deep learning model method.

[0022] Another aim of the invention is to accelerate the process significantly and reduce the workload of experts by means of automatic segmentation algorithms. Thus, it ensures that the tumour regions are determined quickly and efficiently.

[0023] Another aim of the invention is to reduce inconsistencies arising from the subjective interpretations of the expert and provide an objective evaluation by means of computer-assisted segmentation. Thus, it ensures that the tumour boundaries are determined more accurately and consistently.

[0024] Another aim of the invention is to prevent human errors and inaccuracies by means of automatic segmentation algorithms. Thus, it ensures that the tumour regions are determined more accurately and reliably.

[0025] Description of Drawings

[0026] Figure 1 is a drawing showing the flow chart of the Deep Learning Segmentation Model.

[0027] Figure 2 is a drawing showing the flow chart of the Automatic Segmentation Method for Brain Tumours in Computed Tomography.

[0028] Reference Numbers

[0029] 100. Encoder

[0030] 110. Input data acquisition and processing

[0031] 120. Feature extraction

[0032] 130. Transfer of extracted features 200. Decoder

[0033] 210. Reconstruction of segmented output

[0034] 220. Recovery and enhancement

[0035] 300. Pixel-based identification of tumour regions by means of automatic segmentation algorithms

[0036] 400. Data preprocessing by means of a deep learning segmentation model method operating with an encoder-decoder architecture

[0037] 410. Data resampling

[0038] 420. Skull subtraction

[0039] 430. Label map preparation

[0040] 440. Thresholding and cropping

[0041] 500. Training and validation

[0042] 510. Use of stochastic gradient descent

[0043] 520. Updating of parameters in the training data

[0044] 530. Measurement of the difference between the predicted and actual class probabilities by means of weighted cross-entropy loss

[0045] 540. Use of adjustable learning rates

[0046] 600. Accurate identification and labelling of tumour regions and segmentation

[0047] 700. Scaling

[0048] Description of the Invention

[0049] In general terms, the invention is an automatic segmentation method for brain tumours in computed tomography, which ensures accurate and effective segmentation of glioma tumours and generally comprises the process steps of pixel-based identification of tumour regions by means of automatic segmentation algorithms (300), data preprocessing by means of a deep learning segmentation model method operating with an encoder-decoder architecture (400), training and validation (500), accurate identification and labelling of tumour regions for segmentation (600), and scaling (700).

[0050] The invention can be considered as a general process chain. It is a process in which specific steps are combined to process CT images, extract features, perform segmentation, and ultimately achieve automatic segmentation of brain tumours.

[0051] The invention analyses CT images and fully automatically segments glioma tumour diseases. This minimises user intervention and makes the segmentation process simpler and more effective.

[0052] The segmentation results obtained through the use of the invention provide high accuracy and precision. This enables clinical decisions to be made more reliably.

[0053] The technical field to which the invention relates is the automatic detection and segmentation of glioma tumours from Computed Tomography (CT) medical images. Its field of application is Medical Imaging Devices, Radiotherapy, and the Neurology Sector.

[0054] The invention enables an objective evaluation by means of computer-assisted segmentation. Thus, it reduces inconsistencies arising from the subjective interpretations of the expert and ensures that tumour boundaries are determined more accurately and consistently.

[0055] The invention ensures that mutual connections between the encoder (100) and the decoder (200) in the developed deep learning model method help preserve important details during information propagation.

[0056] The invention provides a method developed to achieve faster, more reliable, and more precise results in the segmentation of glioma tumours. This provides significant advantages in areas such as treatment planning, disease monitoring, and research studies in clinical applications. In the deep learning segmentation model of the invention, there are two main components: an encoder and a decoder. The encoder (100) processes input data and extracts meaningful features through convolution layers. These features are then transferred to the decoder (200), where the segmented output is reconstructed using the encoded representation. The decoder (200) utilises sampling techniques to recover spatial information and enhance the segmentation output. Mutual connections between the encoder (100) and the decoder (200) help preserve important details during information propagation. This architecture provides accurate segmentation by using both overall context and fine-grained features.

[0057] The automatic segmentation method for brain tumours in computed tomography comprises the following process steps:

[0058] - Pixel-based identification of tumour regions by means of automatic segmentation algorithms (300)

[0059] - Data preprocessing by means of a deep learning segmentation model method operating with an encoder-decoder architecture (400)

[0060] - Training and validation (500)

[0061] - Accurate identification and labelling of tumour regions for segmentation (600), and

[0062] - Scaling (700).

[0063] Pixel-based identification of tumour regions by means of automatic segmentation algorithms (300) is a process step designed for accurate segmentation of glioma brain tumours from CT medical images, supported by SoftMax and semantic segmentation layers, in order to produce pixel-based classification outputs using the encoder (100)- decoder (200) strategy. By means of advanced neural network algorithms such as residual blocks, skip connections, and concatenation layers, tumour regions are accurately identified on a pixel basis.

[0064] In the data preprocessing step (400) performed by means of a deep learning segmentation model method operating with an encoder (100)-decoder (200) architecture, the “resampling” method is also used to increase the size of the data set. This helps the model to generalise better and to have a more balanced training data set. In the automatic segmentation method for brain tumours in computed tomography, data preprocessing by means of a deep learning segmentation model method operating with an encoder (100)-decoder (200) architecture (400) comprises the following process steps:

[0065] - Data resampling (410),

[0066] - Skull subtraction (420),

[0067] - Label map preparation (430), and

[0068] - Thresholding and cropping (440).

[0069] In the data resampling (410), the images are processed with pixel sizes of 0.5 x 0.5 mm in the x and y dimensions, while the z dimension varies between 3 and 0.5 mm during this process step. This process aims to preserve differences in data augmentation and acquisition angles.

[0070] Skull subtraction (420) is a process carried out by means of a deep learning segmentation model method operating with an encoder (100)-decoder (200) architecture to remove non-brain structures and ensure that the segmentation model focuses only on the relevant anatomical features.

[0071] Label map preparation (430) involves careful definition of tumour core regions for each image by radiologists, and the label maps are prepared by means of a deep learning segmentation model method operating with an encoder (100)-decoder (200) architecture.

[0072] Thresholding and cropping (440) are performed by means of a deep learning segmentation model method operating with an encoder (100)-decoder (200) architecture in order to highlight relevant regions and separate irrelevant areas according to intensity levels, followed by cropping to appropriate dimensions.

[0073] In the automatic segmentation method for brain tumours in computed tomography, the training and validation (500) comprise the following process steps:

[0074] - Use of stochastic gradient descent (510),

[0075] - Updating of parameters in the training data (520), - Measurement of the difference between the predicted and actual class probabilities by means of weighted cross-entropy loss (530),

[0076] - Use of adjustable learning rates (540).

[0077] The invention comprises the use of stochastic gradient descent (510), a widely used iterative optimisation algorithm for training deep learning models. The process includes updating of parameters in the training data (520) calculated with gradients from randomly selected subsets, and measuring the difference between the predicted and actual class probabilities by means of weighted cross-entropy loss (530). This approach enables the deep learning model to learn efficiently from data and improve its performance. The model optimises the deep CNN using a specific learning rate and optimisation algorithm. The learning rate plays a critical role in model training by determining the step size during gradient descent and influencing the convergence speed and model performance. By means of the use of adjustable learning rates (540), the model parameters can be fine-tuned in combination with optimisation algorithms, enabling adaptation to the characteristics of the data and facilitating better training results.

[0078] The use of stochastic gradient descent (510) ensures effective optimisation of model parameters during training by means of the Stochastic Gradient Descent with Momentum (SGDM) optimiser.

[0079] Updating of parameters (520) is performed by continuously updating and improving the parameters using the training data.

[0080] Measurement of the difference between the predicted and actual class probabilities by means of weighted cross-entropy loss (530) ensures that the model effectively learns all classes.

[0081] The use of adjustable learning rates (540) ensures that the learning speed is dynamically adjusted, thereby protecting the model from early convergence and unstable fluctuations.

[0082] For the automatic segmentation method for brain tumours in computed tomography to be implemented, the deep learning segmentation model method additionally comprises the following process steps: - The encoder (100) acquiring and processing input data (110),

[0083] - The encoder (100) extracting meaningful features through convolution layers (120),

[0084] - The encoder (100) transferring the extracted features to the decoder (200) (130),

[0085] - The segmented output being reconstructed in the decoder (200) using the encoded representation (210),

[0086] - The decoder (200) recovering and enhancing spatial information and segmentation output by means of sampling techniques (220).

[0087] In the step of the encoder (100) acquiring and processing input data (110), the encoder (100) receives CT images as input data and processes them. At this stage, the pixel values in the CT image are taken by the model and passed through initial convolution operations.

[0088] In the step of the encoder (100) extracting meaningful features through convolution layers (120), the encoder (100) extracts meaningful features from the input data by means of convolution layers. These features contain structural and morphological information about the brain tumour. The deep learning model ensures that these features are correctly extracted.

[0089] In the step of the encoder (100) transferring the extracted features to the decoder (200) (130), communication between the encoder (100) and the decoder (200) is established, and the features extracted by the encoder (100) are transferred to the decoder (200). At this stage, the deep learning model generates a lower-resolution representation of the image, which is used for subsequent reconstruction.

[0090] In the step of the segmented output being reconstructed in the decoder (200) using the encoded representation (210), the decoder (200) uses the encoded representation to recover the original spatial resolution and reconstruct the segmentation result. At this stage, the boundaries of the brain tumour are accurately determined.

[0091] In the step of the decoder (200) recovering and enhancing spatial information and segmentation output by means of sampling techniques (220), the decoder (200) ensures that the low-resolution representation is restored to its original resolution and that the segmentation quality is improved. At this stage, the spatial information is reconstructed, and the final segmentation result is obtained.

[0092] For the implementation of the data preprocessing step (400) in the automatic segmentation method for brain tumours in computed tomography, which operates by means of a deep learning segmentation model method using an encoder (100)- decoder (200) architecture based on a pretrained ResNet18 encoder (100)-decoder (200) architecture, the following layer steps are included:

[0093] - The input layer receives the input images (“data”).

[0094] - The convolution layer performs convolution with learnable filters and converts the input images into feature maps.

[0095] - The batch normalisation normalises the activations of the convolution layer.

[0096] - The ReLU activation layer applies the ReLU activation function.

[0097] - The max pooling layer helps feature extraction by reducing spatial dimensions and emphasising features.

[0098] - The residual blocks (“Res2a” to “Res5b”) contain multiple convolution layers with skip connections, which increase feature depth while preserving spatial information.

[0099] - In the residual blocks, the encoder-decoder skip connections facilitate information flow and provide better reconstruction by combining features.

[0100] The feature concatenation layers combine low- and high-level features to enable precise segmentation.

[0101] - The SoftMax layer applies the softmax function for pixel classification.

[0102] - The pixel classification layer assigns semantic labels to pixels.

[0103] These layers collectively contribute to feature extraction, information transfer, and reconstruction, thereby enabling effective image segmentation.

[0104] The step of accurate identification and labelling of tumour regions for segmentation (600) ensures precise recognition and labelling of the boundaries of glioma brain tumours. This process enables the model to understand complex spatial relationships within the input data. By increasing the learning capacity of the model, the boundaries of the tumours are accurately determined, and anatomical nuances are distinguished. The scaling step (700) involves dividing the data set by means of a fivefold cross- validation method in order to evaluate the performance of the model within a broader framework and to determine its generalisation capacity. This methodology aims to measure the model’s performance level across different data sets and to examine its scalability. During the cross-validation process, critical performance metrics such as the Dice Coefficient, Jaccard Index (loll), Hausdorff Distance (HD), and Boundary Similarity Coefficient (BSC) are calculated for each fold, and statistical significance tests are applied to these metrics.

Claims

CLAIMS1. An automatic segmentation method for brain tumours in computed tomography, comprising the process steps of- Pixel-based identification of tumour regions by means of automatic segmentation algorithms (300) for producing pixel-based classification outputs by using an encoder (100)-decoder (200) strategy for accurate segmentation of glioma brain tumours from CT medical images,- Data preprocessing by means of a deep learning segmentation model method operating with an encoder (100)-decoder (200) architecture (400), using a “resampling” method to increase the size of the data set,- Training and validation (500) by means of parameter updating, functional adjustments, and learning rates,- Accurate identification and labelling of tumour regions for segmentation (600), by increasing the learning capacity of the model to ensure accurate determination of tumour boundaries and distinction of anatomical nuances, and- Scaling (700) in order to measure the model’s performance level across different data sets and to examine its scalability.

2. The automatic segmentation method for brain tumours in computed tomography according to claim 1 , wherein the process step of data preprocessing (400) comprises the process steps of:- Data resampling (410), in which the images are processed with pixel sizes of 0.5 x 0.5 mm in the x and y dimensions, while the z dimension varies between 3 and 0.5 mm,- Skull subtraction (420) carried out by means of a deep learning segmentation model method operating with an encoder (100)-decoder (200) architecture, in order to remove non-brain structures and ensure that the segmentation model focuses only on the relevant anatomical features,- Label map preparation (430) performed by means of a deep learning segmentation model method operating with an encoder (100)-decoder (200) architecture, and- Thresholding and cropping (440) performed by means of a deep learning segmentation model method operating with an encoder (100)-decoder (200) architecture in order to highlight relevant regions and separate irrelevant areas according to intensity levels, followed by cropping to appropriate dimensions.

3. The automatic segmentation method for brain tumours in computed tomography according to claim 1 , wherein the process step of training and validation (500) comprises the process steps of:- Use of stochastic gradient descent (510) during the training process, by means of the SGDM optimises to ensure effective optimisation of the model parameters,- Updating of parameters in the training data (520) by means of continuous improvement of the parameters using the training data,- Measurement of the difference between the predicted and actual class probabilities by means of weighted cross-entropy loss (530), and- Use of adjustable learning rates (540) in combination with optimisation algorithms to fine-tune the model parameters.

4. The automatic segmentation method for brain tumours in computed tomography according to claim 1 , wherein the deep learning segmentation model method used in the process step of data preprocessing (400) comprises the process steps of:- Acquisition and processing (110) of input data by the encoder (100) by receiving CT images as input data,- Extraction (120) of meaningful features through convolution layers by the encoder (100) by means of convolution operations,- Transfer (130) of the extracted features to the decoder (200) by the encoder (100) by establishing communication between the encoder (100) and the decoder (200),- Reconstruction of the segmented output in the decoder (200) using the encoded representation (210), and- Recovery and enhancement (220) of spatial information and segmentation output by the decoder (200) by means of sampling techniques.

Images