Exploring sequence transformation in magnetic resonance imaging via deep learning using data from a single asymptomatic patient

We investigate the potential for deep learning to create a transfer function from T1 to T2 magnetic resonance imaging sequences using data collected from an asymptomatic patient. Neural networks were trained on images of a human left hand, and then applied to convert T1 images to T2 images for the associated right hand. Analysis showed that the most accurate neural network considered the features in the surrounding ∼1 cm when converting to T2, hence indicating that the neural network was able to identify structural correlations between the sequences. However, some small features measuring <2 mm differed, and grid patterning was evident from the images. While using deep learning for sequence transformations could enable faster processing and diagnosis and in turn reduce patient waiting times, additional work, such as synergising physics-based modelling with neural networks, will likely be required to demonstrate that deep learning can be used to accurately create T2 characteristics from T1 images. In addition, since the present work was conducted using data collected from a single patient, further example datasets collected from patients with a range of different pathologies will be required in order to validate the proposed method.


Introduction
Medical imaging has been a fundamental technique for identifying injuries and health conditions for over 100 years. Key inventions include the invention of x-ray imaging for medical purposes in 1896 by John Hall-Edwards [1,2], pioneering portable radiography vans by Marie Curie in World War I [3], and computed tomography (CT) [4,5]. Other imaging techniques include positron emission tomography scanning (PET) [6], ultrasound [7] and magnetic resonance imaging (MRI) [8,9]. Depending on the imaging technique, the cost of such processes can be high, due to the initial expense of the machinery, cost of maintenance and the expertise needed to run and evaluate results [10]. A single MRI scan in England performed under the National Health Service (NHS) can cost between £53 and £617 (2015-2016) depending on the length (resolution and size of imaging region) and type of scan [11], with larger volumes of imaging and multiple sequences requiring more time and thus expenditure. By way of example, in this work, imaging a human hand via MRI took approximately 4-7 min for each sequence. Therefore, there is much interest in the development of methods for enabling faster turnaround of MRI patients to produce a more efficient and less costly service. In England, 264,520 MRI scans were carried out in the 12 months up to the end of January 2021, with the median waiting time for an MRI scan being approximately 2.5 weeks [12]. Since multiple sequences are often carried out per MRI appointment, methods that can reduce the waiting times via increasing patient throughput, such as by reducing the number of sequences administered out per patient appointment, would be beneficial.
Each medical imaging technique offers distinct advantages and disadvantages, specifically in terms of which features can be identified on the image output. For example, some bone fractures may not be visible on plain radiographs, but might be visible on CT, while different MRI sequences can be utilised to display different tissue characteristics including fat and water content. T1 VIBE (volumetric interpolated breath-hold examination) magnetic resonance images with water excitation show fat in the human body as low signal intensity and water as intermediate signal, whereas T2 SPACE (sampling perfection with application optimized contrasts using different flip angle evolution) images show both fat and water as high signal. As such, work has progressed on being able to synthesise one modality from another, such as MRI to CT scan [54,55]. Additional applications of deep learning for MRI include accelerating magnetic resonance [56], deep learning Bloch equation simulations (DeepBLESS) for rapid and accurate T1 estimation [57], MRI brain extraction in 3D [58], super-resolution musculoskeletal MRI [59], MRI finger printing [60,61], random forest magnetic resonance for MRI synthesis [62] and, most relevant to this work, generation of T2 sequence images from the associated combination of T1 with under-sampled T2 sequence images [63].
As shown by the concept in figure 1, we aimed to use deep learning neural network architecture to generate T2 SPACE sequence images directly from T1 VIBE sequence images by training a neural network on the left hand coronal, sagittal and axial T1 sequence images, to produce an equivalent T2 SPACE sequence output image. The neural network was then tested on the right hand T1 VIBE sequence images to produce T2 SPACE sequence images. Our primary objective for this work was to demonstrate accurate image-to-image generation for MRI sequence images from a single asymptomatic patient, and then establish technical analytical methods for evaluating the effectiveness of the trained network, which could potentially be applied to other studies involving more patients.
A commonly used neural network architecture for paired (input and output are for example, same image slice in space but different sequence) image-to-image transformations is the Pix2Pix cGAN (conditional generative adversarial neural network) model [64] that has a 'U-Net' based architecture for the generator [65], and a convolutional 'PatchGAN' classifier for the discriminator [66], which penalizes structure on the scale of the image patches. Olut et al [67] discuss their results of using Pix2Pix for Magnetic Resonance Angiography (MRA) for generating non-existent MRA from T1-and T2-weighted MRI images, which could be a valuable tool in retrospective subject evaluation of vascular anatomy and related diseases. Zhou et al [68] explore the Figure 1. Schematic of the concept for using deep learning to transform T1 VIBE sequence images into T2 SPACE sequence images, via training on images from the left hand and testing on images from the right hand to generate T2 SPACE sequence images. generation of one MRI sequence from hybrid fusion of two other sequences using Pix2Pix, such as fusing T1and T2-weighted MRI images of the brain to form fluid attenuated inversion recovery (FLAIR)-weighted images. The Pix2Pix model has been used by Shin et al [69] to generate MRI T1-weighted images from segmented labelled images (label-to-MRI), to generate synthetic abnormal MRI images with brain tumours, which could be used for expanding datasets with pathological findings.
An image transformation neural network that utilises unsupervised learning, which has the benefit of not requiring paired training, is CycleGan [70]. This method uses a cycle-consistent adversarial neural network, and has been employed in work for image transformation of MRI T1-weighted sequence images to fractional anisotropy (FA) [71] and computed tomography (CT) images synthesis from MRI images [72,73]. Another type of unsupervised image transformation neural network is UNIT (unsupervised image-to-image translation network) [74], which unlike CycleGan, uses a shared latent space. Implementation of UNIT in medical imaging has been demonstrated for brain image transformation of electroencephalogram (EEG) to functional magnetic resonance imaging (fMRI) images [75], as well as for non-contrast and contrast enhanced CT scans of kidneys [76].
While the work referenced above consisted of using neural networks on data from multiple patients, smaller datasets using single patients have also been investigated in the field of medical imaging. Training on the perfusion MRI data of a single acute stroke patient in order to predict the final infarct of the same patient has been explored by Debs et al [77], while comparisons of deep learning methods trained and tested on single patient ECG for seizure detection were performed by Turner et al [78] and a pilot study of estimating full-dose PET images from low-dose PET images of the whole body were performed by Kaplan et al [79] using training data from one patient and then tested on another patient.
In this work, we initially explore three types of neural network, for generating T2 SPACE sequence images from T1 VIBE sequence images of a single asymptomatic patient, and use the network with the most accurate results for a range of novel digital pathology analysis.

Data collection
The MRI dataset was acquired for the purposes of this study on a single asymptomatic healthy volunteer following informed consent, to establish the utility and efficacy of the method prior to extension to an enlarged data set for any further more in-depth extended study. The MRI examination was performed under standard clinical conditions in accordance with routine safety protocols following a screening questionnaire to exclude contra-indications. Imaging was performed on a Skyra 3.0-T MRI scanner (Siemens, Erlangen, Germany) at the University Hospital Southampton NHS Foundation Trust. Two MRI sequences were obtained in the coronal plane for each hand on a single subject; T1 VIBE (TR/TE msec,13.5/6; flip angle 10 degrees; water excitation; acquisition matrix 512×512; voxel size 0.6 mm×0.6 mm×0.6 mm; acquisition time 4 min 19 seconds for each hand, and T2 SPACE (TR/TE msec,1500/127; acquisition matrix 512×512; voxel size 0.6 mm×0.6 mm×0.6 mm; acquisition time 7 min 5 seconds for each hand. Each set of coronal images were reformatted in the sagittal and axial plane, creating 223 sagittal, 76 coronal and 491 axial image slices. In the analysis section, the reformatted sagittal, coronal and axial image planes correspond to X-Z, Y-Z and X-Y image planes, respectively.

Neural networks
Three types of neural networks were used. These were image-to-image (Pix2Pix), image-to-image-to-image model CycleGan, and unsupervised image-to-image (UNIT). A learning rate of 0.0002 was implemented for all neural networks, which were all trained for 15 epochs for consistency of comparison. The Pix2Pix and CycleGan were trained using an NVIDIA RTX 2080 graphics processing unit (GPU) for a total of 1 h 35 min and 4 h 14 min, respectively, while the UNIT neural network was trained using an NVIDIA QUADRO P6000 and took∼80 h.
Critically, only left hand coronal, sagittal and axial images, each with a resolution of 512×512 pixels, were used in training. After training, each neural network was tested on the right hand T1 images. All 790 images of the left hand (all coronal, sagittal and axial) were used for training the neural network. Data augmentation (i.e. shifting, cropping, rotating and resizing) to increase the amount of training data was not performed in case there was some spatial dependence on the MRI and generation of the T2 images. However, this is something that should be explored in future work through use of a larger sample size, as such analysis may provide insight into the trained neural network, such as the effect of image generation on location or orientation within the 3D volume.
Pix2Pix neural network architecture A diagram of the Pix2Pix neural network architecture used in this work is shown in figure 2. The coloured rectangular boxes represent blocks of multi-channel feature maps, with each map's dimension indicated inside and number of channels indicated below. The U-Net architecture of the generator has a contracting path (yellow boxes) and expansive path (cyan boxes), with skip connections between each centre symmetric layer. The red and blue arrows represent down-sampling and up-sampling convolutions, respectively, while black arrows represent skip connections. Each skip connection concatenates thefeature maps from the expansive path with the equivalent layer feature maps from the contracting path. The contracting path consists of convolutional blocks for down-sampling, in which convolutional filters of size 4×4 with stride 2 are applied to the feature map to double the number of feature channels, followed by a batch-normalisation layer and a rectified linear  unit (ReLU) of 0.2. The expansive path consists of up-sampling, in which 4×4 convolution filters with stride of 2 are applied to the feature map to halve the number of channels, followed by a batch-normalisation layer and a ReLU of 0.2, then a concatenation with the equivalent layer feature maps from the contracting path. The discriminator also has blocks of convolutional layers with a 4×4 convolution filter and stride of 2, followed by a batch normalisation layer and a ReLU of 0.2. An L1 loss (defined as the least absolute deviations) between the generated images and the actual experimental images is produced, such that the L1 loss should be as small as possible for the generator, to allow it to fool the discriminator. The L1 loss of the neural network at the end of training was 0.011, while the GAN loss was 0.69 and the discriminator loss was 0.67. At the start of training, the neuron weightings for the generator and discriminator were randomly initialised. Figure 3 shows a schematic of the Pix2Pix neural network training process, which involves single generator and single discriminator. An actual T1 image was used as input to the generator network (along with a 2D array of noise), which produced a generated T2 image. In line with nomenclature for this field, the experimentally measured images are referred to as actual, and the images produced by the neural network are referred to as generated. During training, the discriminator received either the actual T1 and actual T2, or the actual T1 and generated T2, and had to identify which combination was actual and which was generated. At the same time, the generator was trained to fool the discriminator, by generating images that were visually similar to the real images. The motivation for such adversarial training is to reach a Nash equilibrium [80], where the generated images are indistinguishable from the actual images. At this point, the generator network can be used to convert any T1 image into the associated T2 image.

CycleGan architecture
A diagram of the CycleGan generator and discriminator architecture is shown in figure 4. As per Pix2Pix, a 512×512×3 image is fed into the neural network and down-sampled, in which the channels increase and size of the feature map decreases using blocks of 4×4 convolutional filters with a stride of 2, followed by batch-normalisation and a ReLU activation function. Then the image is passed into 5 residual blocks consisting of 3×3 convolutional filters with a stride of 1, each followed by a batch-normalisation and ReLU. The image is then up-sampled and the number of channels decreased using several blocks consisting of 4×4 convolutional filters with a stride of 2, batch normalisation and a ReLU activation function, apart from the last layer which has a Tanh activation function. The discriminator consists of 4 layer blocks, with convolutional filters with size 4×4 and a stride of 2, with a ReLU following the 1-3 convolutional filters and a batch normalisation following 2-3 convolutional filters.  As seen in figure 5, the CycleGAN neural network consists of two generators and two discriminators, where one generator takes in T1 images and generates T2 images, which are then passed into the corresponding discriminator along with actual T2 images, which tries to correctly classify the images as actual or generated. Likewise, the other generator takes in T2 images and generates T1 images, which are then passed into the corresponding discriminator along with actual T1 images, which tries to correctly classify them as actual or generated.

UNIT architecture
The inputs to the UNIT neural network generator are T1 and T2 images and the output is four images. Two of the output images are transformed images, T1 to T2 and T2 to T1, while the other two are self-reconstructed images, T1 to T1 and T2 to T2. Moreover, the UNIT network consists of a generator that has two input images, one from each domain, T1 and T2, and both of these are fed into their own encoder to give 128×128×256 activations in each block (see figure 6). The outputs of these encoders are concatenated along one dimension to form a 256×128×256 output. Following this, the output is sent into a shared encoder block and then into a shared decoder block, again with 256×128×256 activations. Finally, the output is sent to either a T1 decoder block or a T2 decoder block, where the output has the same dimensions as the input images 512×512×3. Then a T1 discriminator takes in real and generated T1 images and evaluates whether they are realistic, and  showing input T1 image (1st column), actual T2 images (2nd column), Pix2Pix generated T2 images (3rd column), CycleGan generated T2 images (4th column) and UNIT generated T2 images (5th column), for the same coronal view of the centre of the hand. The absolute differences between the generated and actual T2 images are displayed in the 2nd row. likewise a T2 discriminator takes in real and generated T2 images and evaluates whether they are realistic. This is done via a series of convolutional filters and leaky ReLUs [81].
As shown in figure 7, a pair of corresponding images in two different domains, T1-weighted and T2weighted, can be mapped to the same latent code in a shared-latent space, using encoders mapping the images to the latent codes and the decoders mapping the latent codes to the images.

Results and discussion
The three trained neural networks were tested on actual T1 images of the right hand, and the generated T2 images were compared to the actual T2 experimental images from the single asymptomatic patient. Figure 8 shows a 512×512 1 channel 8-bit input T1 image (first column), actual T2 image (second column), Pix2Pix generated T2 image (third column), CycleGan generated T2 image (fourth column) and UNIT generated T2 image (fifth column) of the same coronal view of the centre of the hand. Since the pixel intensity of an MRI sequence image is related to specific characteristics (i.e. of bone, fat etc), normalising the image (such as to the maximum value) would change the intensities and thus characteristics displayed. The absolute difference between the T2 generated images and the actual T2 images are displayed below their respective generated image (one minus the other), such that the higher intensity (whiter value) signal corresponds to greater difference in pixel intensity value.
It is evident upon looking at figure 8 that the Pix2Pix, under the same learning rate and number of epochs, has more accurately generated the actual T2 image, compared with the CycleGan and UNIT neural networks. Indeed, the characteristics and associated colour and shape seem to be preserved. This is perhaps due to the multiple concatenation connections and greater depth of the Pix2Pix neural network. The absolute difference image between the generated T2 image for Pix2Pix is visibly darker compared with the CycleGan and UNIT absolute difference images, indicating a greater accuracy of image generation for Pix2Pix. Error analysis on the images also indicates this, and is detailed below.
The normalised root mean square error (NRMSE) was calculated by taking the mean of the squared difference between the intensity value of each pixel in the generated image (0-255 intensity range) and that of the actual experimental image (0-255 intensity range), where N is the number of data points (pixels), G i is the generated pixel value and I i is the actual pixel value, with I imax being the maximum pixel value and I imin being the minimum pixel value of the actual image. The lower the value of NMRSE, the smaller the difference between the generated and actual images. The mean NRMSE and standard deviation for all the generated images is 0.0396±0.0175 for Pix2Pix, 0.0844±0.0596 for CycleGan, and 0.0473±0.0207 for UNIT. To further quantify the performance of the image generation by the Pix2Pix neural network, we determine the Peak Signal-to-Noise Ratio (PSNR) of all the images generated, defined as, In addition, we have also calculated the Structural Similarity Index Measurement (SSIM) between the generated and the actual images, which assesses the visual impact of image luminance, contrast and structure, defined as, where μ I is the average of I, μ G is the average of G, σ I 2 is the variance of I, σ G 2 is the variance of G, σ IG is the covariance of I and G, C 1 =(0.01 L) 2 and C 2 =(0.03 L) 2 , such that L is the dynamic range of the pixel values.
These results are presented in table 1 below. The mean SSIM and standard deviation for all the generated images is 0.9676±0.0359 for Pix2Pix, 0.3921±0.3612 for CycleGan, and 0.8230±0.0650 for UNIT.
The lower the NRMSE, the higher the PSNR value and the higher the SSIM value (max 1), the higher accuracy of generated images compared with the actual experimental images. Visually, and as shown by the error numbers, the Pix2Pix is clearly the most accurate neural network trained under the conditions (learning rate and epochs) described in this work.

Analysis
Since Pix2Pix produced the most accurate image generation, the results were further explored. Additional examples of the Pix2Pix neural network image generation is shown in figure 9, which presents the input T1 image (first column), actual T2 image (second column), generated T2 image (third column), and absolute difference (forth column) for (a) coronal view at the centre of the hand, (b) sagittal view down the middle finger and below, and (c) axial view of the wrist. The fourth column shows the absolute difference between the actual and generated images.
The neural network was able to generate visually similar images to the actual sequence type. As shown in figure 9, muscle on the generated T2 images is of similar intermediate signal intensity to the actual T2 images, Figure 9. Capability of the trained neural network for generating a T2 image of the right hand, showing T1 images (1st column), actual T2 images (2nd column), generated T2 images (3rd column) and absolute difference between actual and generated (4th column), for (a) coronal, (b) sagittal (along the middle finger) and (c) axial planes. and bone marrow is of similar high signal. Specific features present in the actual images, such as localised regions of high signal are also present in the actual and generated images, for example at the 5th carpometacarpal joint between the hamate and the base of the 5th metacarpal in figure 9(a). As shown in table 2, the NRMSE between the generated and the actual T2 image for coronal, sagittal and axial images shown in figure 9 is 0.0650±0.0188, 0.0315±0.0081 and0.0440±0.0191, respectively. The mean PSNR and standard deviation for the images in the coronal plane, sagittal plane and the axial plane is 26.4±2.4 dB, 34.9±5.2 dB, and 48.8±15.7 dB, respectively. The values of the SSIM are also shown in table 2, with a mean value of 0.8448±0.0212 for the images in the coronal plane, 0.9759±0.0053 for the images in the sagittal plane, 0.9768±0.0192 for the images in the axial plane. The differences in the NRMSE, PSNR and SSIM between the 3 planes can be understood due to the area of the image in which the features of the Figure 10. Actual T2 image (left), generated T2 image (middle) and magnified generated T2 image (right), for the metacarpal bone of the thumb. An artefactual grid pattern can be seen on the generated image.  hand are present, with coronal plane containing more hand structure than the sagittal and the axial. It is important to realise that actual T2 and generated T2 images would not be expected to be identical, due to small movements in the patient's hand position and orientation between sequence measurements. In addition, disparities between the patient's left and right hands, such as different degrees of muscle mass or bone density, may also have resulted in an imperfect prediction. Additional training data could help reduce such an effect.
Since this work has only involved data from a single asymptomatic patient, in this section we focus on demonstrating a wide variety of analytical methods, which could be applied to larger studies and even point of care research. More specifically, we perform a range of digital pathology techniques, such as image cross-section analysis and sectional image generation to understand the resolution limitations of the generated images and the amount of information required from the input image for successful output image generation. On closer inspection of the images, it is clear that although there is similarity in the signal intensity and shape of the actual and generated images, as visualised in figure 10, small structures can be different in shape (see for example, labelled regions on images). It is also evident that the generated images contain a grid pattern, most likely a result of the discriminator neural network resolution. It is necessary for both faults to be corrected for future work if neural network transformations are to be applied to clinical practice. Further to this, on comparing the signal intensity of slices through the actual and generated images of the metacarpal bone of the thumb, as shown in figure 11, it is evident that the generated signal intensities are generally higher, and the peaks generally broader, as compared to the actual T2 images.
To allow a more detailed analysis of the neural network effectiveness, a small region from the axial plane is presented in figure 12, which shows (a) actual T1 image, (b) actual T2 image, and (c) generated T2 image. Since the training was carried out on the left hand and testing was carried out on the right hand, the network was not purely memorising features. Of interest, therefore, is the numerical relationship between the actual T1 image and actual T2 image, compared with the actual T1 image and generated T2 image. This analysis is shown in figures 12(e) and (f) respectively, which show scatter plots showing the signal intensity values of T1 image pixels against the intensity values of the associated T2 image pixels. These scatter plots show that the transformation of intensity from T1 to T2 is a one-to-many mapping, but that the neural network has indeed, to a degree, replicated this transformation. The images imply that the neural network is processing the structure and intensity when predicting the associated T2 image, rather than simply changing the intensity of each image pixel. As a further comparison, figure 12(d) shows the result of computing T2 via a one-to-one transformation, formed by taking the most probable T2 image signal intensity value for each T1 image pixel from figure 12(e). The fact that this simple computation produces an output that bears almost zero resemblance to the actual T2 is further evidence that the neural network has indeed learned to recognise features within the images.
To quantify the sizes of features used by the neural network when making a T2 prediction, the effect of adjacent pixels on the generated T2 images was explored. This was achieved by analysing the predicted T2 output whilst restricting the amount of information present in the T1 image, through setting signal intensity values of pixels in specific regions of the T1 image to zero. This method can be observed in figures 13(a)-(d), which shows the T1 image with varying amounts of information present, via setting the width of the observable window from 1-pixel to 100-pixels wide. Figures 13(e)-(h) show the associated neural network predictions for these restricted images, which clearly show that as the T1 image is uncovered, the generated T2 image similarly uncovers. Figure 13(i) shows the intensity values of three pixels (indicated by the red, green, and blue crosses in figure 13(e)), as a function of the width of the observable window. On the plot, the horizontal dashed lines show the intensity value at the same pixel positions on the actual T2 image. The fact that the intensity values on the generated T2 image do not match the actual value until approximately an observable window of approximately 20 pixels, implies that the neural network requires at least 20 pixels of information to make a strong prediction. Based on the scales in the images, 20 pixels corresponds to 1.01 cm, and hence this figure provides evidence that the neural network considers the structural information present in the surrounding∼1 cm when transforming a T1 image into a T2 image.
Finally, as shown in figure 14, the neural network was applied to all T1 image slices in the axial plane to generate T2 images for the entire volume. Figure 14(a) shows eight axial planes of generated T2, whilst figure 14(b) shows the result of combining all axial images together to produce a complete and generated T2 image volume.
A richer data collection from a variety of individuals, which includes abnormalities, along with exploring open datasets, such as the multimodal Brain Tumour Image Segmentation Benchmark (BRATS) [82] and OpenfMRI [83], will enable rigorous investigation of such methods. More specifically, our previous work on Pix2Pix for accurate image generation of (∼5-pixel resolution) has involved approximately 100 different shapes, and so increasing the dataset to include, not just different image planes but additional people/objects would be beneficial. In addition, we found that variation is as important as number. Indeed, we have often used 1000s of randomly created shapes in order to cover as many possible angles/sizes and variations as possible, and this is where phantoms could be useful. Since the area of the reconstruction was limited to∼1 cm, it could be useful to train using phantoms of smaller than that size in order to potentially allow the neural network to generate accurate characterisations on a smaller scale. Since we found the surrounding structure to be important to the image generation, this could further aid in examination of any links between tissue boundaries and abnormalities, or other cues of surrounding areas. Moreover, perhaps moving to 3D data and using a 3D U-net architecture [84] could aid in more accurate characteristic generation if there is a reliance on surrounding structure for characteristic generation accuracy, since this would add in an extra dimensional constraint to the training and learning. This could be done along with examination of the convolutional filters and their activations, to allow for greater understanding of the transformation function and quantifying the neural networks capabilities. Determining the limitations could unlock methods for improving the image generation, either through additional data or via physics-guided networks, where one incorporates physics-based modelling [85]. This should ideally be done using 3D image volume data, and where possible similar to previous image segmentation approaches [86], in which one could add physics branches to the architecture that consists of the MRI physics parameters or physics-based static-equation sequence simulations to help the networks encode a Figure 14. Generation of a 3D MRI T2 volume, generated from a neural network directly from T1 images, showing (a) eight generated T2 axial image slices along the hand and (b) a combination of all generated axial T2 image slices with a false colour map (blue indicates low signal intensity, yellow indicates high signal intensity) to enhance visual clarity. more accurate representation of the physics underlying MRI. Furthermore, one could use a CycleGAN [70] network to cycle between the physics-guided neural network generated images and the actual images, in which one varies the input parameters and thus the T1-weighted and the T2-weighted sequence images. This could perhaps create a neural network that has an understanding of the underlying MRI physics and produce more accurate characteristic generation.

Conclusion
To conclude, using data from a single asymptomatic patient, we have utilised neural networks to generate T2 SPACE sequence images from T1 VIBE images and subsequently performed a range of pathologies on the results to aid in understanding the capabilities of the deep learning methodology. Analysis showed that the Pix2Pix neural network was able to generate the similar signal through consideration of the surrounding∼1 cm of image information. Specifically, regions of high signal intensity were correctly observed in a generated T2 image, such as between the hamate and the base of the 5 th metacarpal. However, some signal intensities were 50% less than the actual and the resolution of some structures in the generated images was not as high as those in the actual. In addition, there were regions of artefactual grid patterning on the images. To resolve these issues and improve the transformation accuracy, a larger and more varied training dataset could be used, i.e. images from more than one person and area of the body. Additional work will be needed to prove that the neural network can accurately create T2 characteristics from T1 images, such as via synergising physics-based modelling with deep learning to understand and enhance the MRI technique. Also, other T2 sequences such as the STIR (short tau inversion recovery) sequence that nulls the signal from fat could be investigated, since for clinical diagnosis alternative weighted sequences would allow information on the characteristics involved, and thus assist in further quantification of the accuracy of the generated images. Furthermore, imaging could be carried out on parts of the body in which there are known abnormalities, such that they are not identified in one weighted type (e.g. T1-VIBE), but present in another (e.g. T2-STIR), to further evaluate the neural network. This initial work could be the basis for faster processing and diagnosis, and in turn, reduce the patient waiting times, in addition to being a useful mathematical tool for assisting in fundamental research of magnetic resonance imaging.