Three dimensional convolutional neural network-based automated detection of midline shift in traumatic brain injury cases from head computed tomography scans

Data acquisition

Head CT scans were prospectively collected from the Department of Neurosurgery, All India Institute of Medical Sciences, New Delhi, for the study. The study received ethics approval from the Institute’s Ethics and Scientific Committee. The data collection spanned a period of nine months from January 2021 to September 2021 and included consecutive head CT scans of patients coming to the neurosurgery department. The head CT scans were obtained from two different machines, namely Siemens (Germany) and GE Healthcare (Illinois, United States of America). At the hospital, all head CT scans were stored in the standard Digital Imaging and Communications in Medicine format, which includes text information embedded within the images. Accessing and displaying the CT images was facilitated through the Oviyam Picture Archiving and Communication System (PACS) hardware. The head CT scans along with their corresponding clinical data were retrieved from the Oviyam PACS system. Consecutive clinical data including demographic details and Glasgow coma score were downloaded from the clinical patient record system. The radiologist’s report of the CT scans, used as the gold standard for evaluation, was also included in the metadata collected alongside the head CT scans. The local Picture Archiving and Communication System (PACS) servers were the source of data extraction and storage for this study. To ensure patient privacy and compliance with regulations, all data underwent anonymization following the Health Insurance Portability and Accountability Act guidelines.

Head CT dataset

Upon acquisition, the collected data underwent comprehensive analysis to assess various aspects of the head CT scans. Among the considerations, the slice thickness was examined, ranging from 0.625 mm to 5 mm. For inclusion in the study, only scans with a slice thickness of 1 mm and a 30 s kernel were considered. In addition, head CT scans with discernible artifacts were excluded from the dataset to ensure data quality. To differentiate true MLS cases from those mimicking normal cases or artifacts, a knowledge-based classification system was employed. This system incorporated rules that utilized quantified imaging features and anatomical information. The dataset was divided into training, validation, and testing classes facilitating the development and evaluation of the automated detection system. The training class encompassed a total of 156 head CT scans consisting of 78 normal cases and 78 cases with MLS. The testing class, used for evaluating the system’s performance, comprised 10 normal cases and 10 MLS cases. Consequently, the dataset contained a total of 176 head CT scans.

Technological approaches for automated detection

The chosen technique for automated detection in this study is the three dimensional (3D) convolutional neural network (CNN). The architecture of the 3D CNN comprises convolutional layers, each consisting of fixed-size convolutional filters responsible for extracting relevant features from the input images. Subsequently, these features are spatially reduced and accumulated using pooling layers, employing either max pooling or average pooling techniques. The extracted features are then passed through fully connected layers to reach the output units of the network. The utilization of fully connected layers, along with appropriate activation functions, enables the classification of inputs based on the reduced set of feature vectors.^[4]

Developing the model considered the quantity and quality of the available data. During the training phase, the model was monitored using validation metrics and employed the “save best model” option to optimize performance. The pre-processing, hyperparameters, and data augmentation techniques were applied consistently across the training of the 3D CNN models. However, variations were introduced in the training modules to achieve the best results. Below are the details of the specific model configuration that attained the highest performance outcomes during experimentation.

Implementation and model configuration

Architecture of the 3D CNN Model

The architecture of the 3D CNN model is depicted in Figure 1. It comprises four blocks followed by a global average pooling layer, two fully connected dense layers, and a dropout layer. The input to the CNN architecture consists of head CT scans with dimensions (128, 128, 64) and one channel. The architecture is composed of four blocks each containing 3D convolutional layers with subsequent ReLU activation functions. In addition, each block includes a 3D MaxPooling layer and a batch normalization layer to enhance the learning process. The dimensions of the input are progressively reduced through the four blocks. The first block reduces the dimensions to (63, 63, 31, 64), the second block further reduces to (30, 30, 14, 64), the third block reduces to (14, 14, 6, 128), and the fourth block further reduces to (6, 6, 2, 256). Following the final block, a global average pooling layer is applied to generate a vector. This vector is then passed through a fully connected layer, with a dropout layer applied subsequently to mitigate overfitting. The output layer consists of a single neuron producing a binary classification output (1 or 0) indicating the presence or absence of hemorrhage in the head CT scan. The performance of the algorithm was evaluated using accuracy and loss values.

Close

Figure 1:

Proposed 3D convolutional neural network architecture for midline shift. CNN: Convolutional neural network.

Export to PPT

Model analysis and performance

Our study demonstrates that deep neural networks trained on clinical imaging datasets can effectively detect critical radiological conditions such as MLS with reasonable accuracy. The screening tool underwent testing on 20 cases to assess its performance. The training model exhibited the ability to detect 7 MLS cases out of the 10 true MLS cases and correctly identified 4 non-MLS cases out of the 10 true non-MLS cases resulting in an overall accuracy of 55%. In addition, the model showed a specificity of 70% and a sensitivity of 40%. The training plot graph, depicting the accuracy and loss of the model during the training process, is illustrated in Figure 2. The recorded values for loss and accuracy were 0.4682 and 0.7133, respectively while the validation set yielded a loss of 0.4628 and an accuracy of 0.7368. The confusion matrix for the best-performing training plot providing a visual representation of the model’s performance is presented in the Figure 3. The robust performance of our deep learning algorithms indicates their promise in potentially enhancing the speed and accuracy of MLS detection in head CT scans.

Close

Figure 2:

Training plot graph showing model’s accuracy and loss. train: Training dataset, val: Validation dataset, acc: Accuracy.

Export to PPT

Close

Figure 3:

Graphical representation to show confusion matrix for the best preforming training plot. MLS: Midline shift.

Export to PPT

The symmetry-based approach

This approach is based on the concept of symmetry around the brain’s midline. It does not require the recognition of specific anatomical landmarks. Instead, it seeks to identify a curve that connects all displaced and deformed structures known as the “deformed midline” (DML).^[4] Liu et al.^[5] developed the hemorrhage-MLS heuristic model to explore the relationship between intracranial hemorrhage and MLS. The model utilized 11 CT images and 423 midline points employing 10-fold cross-validation. However, clear results were not provided. A ventricular geometric pattern-related approach was done using anatomical information from 170 CT sets to identify the ideal midline and the MLS.^[13] Liao et al. proposed the “skull axis” method for MLS measurement on CT slices at the level of the foramen of Monro. They tested the algorithm on 81 clinical cases demonstrating moderate accuracy in detecting MLS and lower accuracy for large MLSs (>5 mm) with spontaneous intracerebral hemorrhage. However, their approach had limitations including manual slice selection and limited performance in severe TBI cases.^[7] On the lines of symmetry-based approach, a shape-matching technique was used for MLS identification aiming to correlate MLS with intracranial pressure (ICP) levels. The method was evaluated on 391 slices from 17 TBI patients yielding errors of approximately 1 mm for the ideal midline and <2.25 mm for the DML in over 80% of cases. However, difficulties were encountered when ventricles could not be identified due to significant brain deformation. The study achieved an accuracy of approximately 70% with a sensitivity of about 65% and specificity of about 73%.^[12]

Some researchers faced challenges in estimating larger MLS and focused on addressing this issue. One such paper developed a computer-based approach using MATLAB algorithms to detect MLS larger than 5 mm. The software was validated on over 200 patients suspected of acute TBI demonstrating high sensitivity (98%) for detecting radiological signs of acute TBI. The method achieved a sensitivity of 92% and specificity of 85% for detecting MLS larger than 5 mm. However, due to limited patients with such findings and additional false-positive results, the positive prediction rate of their MLS detection method was only 70%.^[14]

The landmark-based approach

In landmark-based algorithms, the focus is on recognizing specific structures, typically frontal horns of the lateral ventricles, septum pellucidum, third ventricle, etc. These structures serve as anatomical markers for MLS measurement. Among the identified landmarks, the most suitable ones are chosen to construct the DML, which is then compared to the ideal midline for quantifying MLS. This approach allows for targeted analysis and measurement of MLSs based on anatomical markers within the ventricular regions.^[6,14] One of such landmark-based approach developed a fully automatic tool, which showed a strong correlation with manual drawing, detecting both small MLSs (<2 mm) and large MLSs (>10 mm). This study demonstrated good sensitivity (84.6–91.7%) and specificity (80–97.4%) for detecting MLSs of 2 mm and 5 mm > 10 mm.^[3] A quantification tool, called Icobrain, is based on 2D U-Net and can identify and measure the MLS. The study conducted for developing Icobrain included a dataset of 38 images and achieved a median absolute difference in MLS of 0.86 mm and a classification accuracy of 0.89 for MLS detection at 5 mm threshold.^[15] A CNN-based model was employed to estimate the extent of MLS in TBI patients.^[16] In 2017, a study introduced an MLS measurement method with an average processing time of approximately 10 s. The automated system exhibited an overall accuracy of 90.24%, which further improved to 92.68% with manual calibrations. This method also reported an accuracy of 0.90 for detecting MLS larger than 5 mm in a study involving 43 subjects.^[17]

One approach proposed MLS measurement by recognizing the septum pellucidum and tested the system on images from 96 patients. The algorithm demonstrated accuracy in measuring MLS up to 30 mm with a mean difference of 0.23 ± 0.52 mm compared to manual measurements in 78 cases. This approach achieved a sensitivity of 94%, specificity of 100%, and a positive predictive value of 100% for MLS >5 mm. However, accurately measuring large MLS with significant hematoma remained challenging.^[18,19] In another study, Gaussian mixture clustering was applied to an experimental dataset with 565 patients for automated MLS detection and quantification. The method achieved a maximum distance error of 4.7 ± 5.1 mm, with over 100 patients having MLS larger than 5 mm.^[5]

Chilamkurthy et al.^[20] conducted MLS detection on two datasets: Qure25k (25,000 images, area under the curve [AUC] 0.93) and CQ500 (500 images, AUC 0.97). Their method achieved an average sensitivity of 0.89 for MLS detection (>5 mm). On the other hand, Nguyen et al.^[8] proposed a landmark-based system with two modules for MLS detection and quantification using CQ500 dataset achieving a total accuracy of 0.89. Most research efforts focus on MLS quantification, some primarily emphasize measuring the shift rather than its detection. In contrast, our work is specifically developed with the purpose of MLS detection contributing to the advancement of automated MLS detection and quantification in TBI cases. The developed 3D CNN model represents a significant step toward improving MLS detection accuracy and has the potential to enhance diagnostic efficiency in TBI management.

Future directions

Identification of MLS in an emergency setting is a strenuous job. As a radiologist, one must read all the CT scans of a lot of patients and look out for abnormality in all the normal cases.^[3] The assessment of secondary injuries post-TBI such as MLS and ICP, holds significant clinical relevance for patient management.^[15] Among the various imaging modalities, CT of the brain is commonly performed and interpreted by radiology trainees.^[5,14] Previous studies have consistently demonstrated a strong correlation between a considerable MLS observed on CT scans and adverse outcomes including disability or mortality in TBI cases.^[13,21] One crucial measurement utilized in the assessment of brain symmetry changes is the MLS, which serves as an essential indicator of pathological severity. Severe brain trauma leads to brain swelling causing imbalanced pressures between the left and right hemispheres. This pressure imbalance leads to the shift and deformation of the ideal midline into a curve causing the MLS.^[5,13] A MLS exceeding 5 mm on initial brain CT images is considered significant and predictive of poor neurological outcomes.^[3,17,21] While the current work focuses on MLS detection, it does not quantify MLS. To further enhance the outcomes of this study, future directions involve expanding the dataset and continuing to train the 3D CNN model. Including data with various slice thicknesses, kernels, and larger datasets will enhance the robustness of the model and improve overall accuracy. With broader dataset incorporation, the automated MLS detection system can become a valuable triaging tool or an immediate notifier for patients with critical findings, particularly in emergency settings upon acquiring a head CT scan. The proposed system holds promise for facilitating teaching and research endeavors and streamlining the search through large datasets of CT scans. However, to fully harness the potential of this model, additional studies are required to address limitations and validate performance across diverse patient populations and different scanning equipment.