Ethical approval

The Internal Review Board at Boston Children’s Hospital approved this study (IRB-P00029015) and determined that it met the regulatory requirements to obtain a waiver of inform consent/authorization from research subjects as this study was a secondary use of already existing data collected primarily for clinical reasons.

Study design

This study applies CNNs, a type of neuronal network developed for computer vision, to detect tubers in brain MRI images. For a more in-depth overview of CNNs, transfer learning, data augmentation, and visualization techniques relevant to this article, please see Supplementary Methods at: https://ivansanchezfernandez.github.io/TSC_supplementary_methods/. All supplementary files at dx.doi.org/10.17504/protocols.io.bdt3i6qn

Patients

Our population of interest consists of children and adolescents with TSC and tubers visible on their MRI. Our representative sample consisted of patients with TSC followed at the Multidisciplinary Tuberous Sclerosis Clinic and controls with normal MRI from Boston Children’s Hospital, a tertiary pediatric center with a heterogeneous sample of patients with TSC of a wide variety of ages, severities, and comorbidities. The inclusion criteria for patients with TSC were: 1) patients with a confirmed diagnosis of TSC following the 2012 International Tuberous Sclerosis Complex Consensus Conference Diagnostic Criteria [12], 2) who had at least one brain MRI at Boston Children’s Hospital, 3) who had both T2-axial and FLAIR-axial sequences available, and 4) who had tubers detected clearly on MRI as per the radiology report during routine clinical care. The inclusion criteria for the control group were: 1) patients who had at least one brain MRI at Boston Children’s Hospital, 2) the MRI was interpreted by the neuroradiologist as normal or with non-specific findings during routine clinical care. Common indications for obtaining an MRI in controls were headache, concussion, non-syndromic mild developmental delay, and treatment-responsive epilepsy. When several brain MRIs were available per patient, we selected the most recent MRI to minimize the decreased lesion contrast associated with immature myelination. When a patient developed hydrocephalus or underwent shunt placement or epilepsy surgery we selected the last brain MRI before such interventions.

MRI findings in TSC

Cortical tubers are not the only findings of TSC in brain MRI [5]. Subependymal nodules and white matter radial migration lines were also present in some of our MRI images. However, since tubers were present in all images labeled as TSC and not present in any of the images labeled as controls, we expected the CNN to detect tubers as the pattern that differentiated TSC patients from controls. On brain MRI, tubers are moderately well-circumscribed areas of increased signal intensity on T2-weighted and fluid-attenuated inversion recovery (FLAIR) images. The cortex overlying the tuber may have features of malformation, and the gray-white matter differentiation is reduced [5].

MRI sequences, image labeling, and division into training, validation, and testing

We only selected two-dimensional axial T2 and FLAIR sequences for both patients with TSC and controls. MRIs were collected for clinical reasons. The objective of the study was to develop a deep learning algorithm to detect tubers in MRI images (not to identify from patients with TSC from any image, as not all individual slices from patients with TSC contain lesions). For the MRI of each patient with TSC, a pediatric neurologist (ISF) selected several MRI slices in axial planes with obvious tubers in them. For the controls, the same pediatric neurologist selected MRI slices at approximately the same level in the brain that for patients with TSC. The images of 138 patients (69 TSC and 69 controls) were used for training the model, of 40 patients (20 TSC and 20 controls) were used to validate the model, and of 50 patients (25 TSC and 25 controls) were used to test the model. There was no patient overlap between the training, validation, and test sets. We developed and validated the model based only on labels derived from routine clinical radiology reports to demonstrate that these models can be developed with data generated from routine clinical care. The labels (TSC or control) for the images in the testing set were independently confirmed by a clinical neuroradiologist (EY).

Minimizing overfitting

CNNs are complex mathematical functions with a huge number of parameters, which allow them to fit well complex datasets but, at the same time, makes them prone to fit the data too well with poor generalization. We used several techniques to minimize overfitting:1) keeping training, validation, and test sets completely independent of each other (held-out cross-validation, as is the standard approach in CNN); the test set consist of data the final deep learning model was never exposed to before and, therefore, it is a good evaluation of how well the final deep learning model will do on new data, 2) using random noise, batch normalization, dropout, and global average pooling in the CNN architectures, and 3) using data augmentation (https://ivansanchezfernandez.github.io/TSC_supplementary_methods/).

Data augmentation

We augmented the training set by creating approximately 4 copies of each original image by randomly allowing shifts, horizontal flips, and rotations. Data augmentation of the training set is a standard training approach in CNN which makes CNNs more robust to the essential features of each class and less sensitive to particularities of the individual images used for CNN training such as laterality (left or right), location of the tuber in the image (for example, upper part of the image or lower part of the image), rotation, etc. [13]. In addition, data augmentation of the training set has a crucial role to obtain good CNN performance using relatively small datasets like the present one, effectively increasing the size of the training set for the CNN to learn without increasing the number of subjects [13]. We augmented our training data with Image data generator from Keras [14] using the following parameters: rotation range: 30 degrees, width and height shift range: 0.1, shear range: 0.2, zoom range: 0.25, we allowed horizontal flip; and we used nearest neighbors as the fill mode. The validation and test sets were not augmented.

Variables

The primary outcome was the accuracy of the deep learning algorithm to correctly classify the MRI image as TSC or control in the test set (using the best performing CNN model in the validation set, as explained below). The secondary outcome was the ability of the deep learning algorithm to detect tubers within each image in the test set.

Model development

CNNs process data in the form of arrays. In the case of two-dimensional images, each array is three-dimensional with two dimensions for width and height and one dimension for the three color channels (red, green, and blue), even when the MRI image is grayscale. Each numerical value in the three-dimensional array represents the pixel intensity in the three color channels to construct the grayscale MRI image. We downscaled the input size of the images to 224x224x3 and normalized pixel values between 0 and 1, which are standard transformations for input data used in CNNs. A CNN has an architecture with several layers so that the inputted image is transformed in several steps to eventually yield a class prediction [15]. The main types of layers are convolutional layers, pooling layers, and dense (fully-connected) layers. Convolutional layers “scan” different areas of the image trying to find spatial patterns, for example, edges. Pooling layers ensure that the local conjunction of features from the previous layers are detected regardless of their location in the image (translation and rotation invariance). The convolutional and pooling layers automatically perform most of the feature extraction and feature transformation and, towards the end of the CNN architecture, the image is flattened to a one-dimensional vector. Afterward, fully-connected layers try to translate the feature vector into probabilities of the original image belonging to one of the classes [15]. Each CNN has thousands to millions of parameters to be tuned iteratively. These parameters have no direct interpretability, they are just weights in a complex mathematical function with no intuitive interpretation. During the forward pass the initial weights yield predictions. These predictions are compared with the ground-truth image labels and an error measure is calculated. The error is “backpropagated” with partial derivatives, so that the “responsibility” of each parameter in the error is calculated and the parameter is slightly changed in the direction that reduces that error. Through multiple forward passes and backpropagation steps, the parameters are tuned to optimize classification [15]. We tried three different CNN architectures: 1) Tuberous sclerosis complex convolutional neural network (TSCCNN), a relatively simple architecture that we developed with 4 blocks, each of them consisting of several convolutional layers followed by a pooling layer, and a final block of fully-connected layers (https://ivansanchezfernandez.github.io/TSC_supplementary_methods/), 2) InceptionV3, a popular architecture within the family of CNNs that parallelize computations in a split-transform-merge approach to increase depth and improve accuracy while keeping computations efficient [16, 17], and 3) ResNet50, a popular architecture within the family of residual CNNs that allow very deep CNNs by using blocks of layers that behave like relatively shallow classifiers and work together as an ensemble to produce a very good classifier [18–20]. The initial weights for TSCCNN were random weights, while the initial weights for InceptionV3 and ResNet50 were the transfer learning weights used during the ImageNet competition [21]. Transfer learning is the improvement of learning in a new task through the transfer of knowledge from a related task that has been already been learned and can potentially help performance in relatively small datasets [21]. Initialization weights are tuned in each step of backpropagation during training. However, initialization weights already trained to identify unrelated items in different images are easier to train than initial random weights because image recognition largely implies detection of edges and combination of these edges [21]. For all CNNs, we used Adam [22] as an optimizer with a learning rate of 0.00025, we used binary cross-entropy as loss function, and a batch size of 64 with 100 epochs.

Model validation

The validation set evaluates how well the CNNs would generalize to images they were not trained on. We selected the most generalizable model as the one with the lowest binary cross-entropy loss in the validation set. Binary cross-entropy loss is a more granular and more stable measure of generalizability than accuracy, since it is a continuous measure as opposed to a dichotomous measure for classification.

Model testing

While the purpose of model validation is to select the potentially most generalizable CNN, the purpose of model testing is actually testing generalizability in a completely new set of images: the test set. The primary outcome of classification accuracy was calculated as the proportion of images in the test set labeled correctly by the model. There was no patient overlap between the training, validation, and test sets, therefore the images in the test set came from patients who were not used in the training set or validation set and to which the CNNs were never exposed before.

Model development versus clinical practice: Clinical cases

The development of a CNN entails obtaining observations (in our study, MRI images) and their labels (in our study, images containing tubers or images derived from a normal MRI) by a human to first train, then validate, and finally test the model. However, once the model has been trained, validated, and tested, its use in clinical practice is to assist a clinician in the recognition of new observations as belonging to one of the labels (diagnoses). In our study, that would mean recognizing whether an MRI image has tubers in it or not. As an illustration on how we envision the clinical use of our deep learning algorithm, we further evaluated the selected deep learning architecture in a series of 259 consecutive MRI T2 and FLAIR axial images from 6 additional patients with TSC. These images were not used in training, validation, or testing because of challenging characteristics, such as extremely subtle tubers or limited myelination.

Model visualization

Machine learning can sometimes operate as a “black box” where it is not possible to know which features are used in the classification process. However, model visualization in CNN helps clarify the most relevant features used for classification. For example, a deep learning model used to diagnose disease from chest X-rays classified images as having pneumothorax because of the presence of a chest tube in the image [23]. Similarly, some deep learning models focused on areas in the image that indicated the origin of the image as hospital-based (portable X-ray) versus images obtained in the doctor’s office to classify chest X-rays as normal or pathological [24]. Intuitive and graphical indications on how the CNN is making decisions helps users gain confidence in the model. Thus, to identify potential correct classifications based on incorrect features and to gain some intuition into the classification process, we identified image pixels most relevant for classification with gradient-weighted class activation maps (Grad-CAM) and saliency maps. Grad-CAM uses the gradient of the output category (TSC or control) to the last convolutional layer (the last layer with spatial information) to provide a coarse localization of the areas of most interest for classification (where the CNN is “preferentially looking at for making decisions”) [25]. The objective of Grad-CAM and saliency maps is to make sure the CNN is classifying MRI slices based on what it is supposed to. The objective of this project was not to segment the patterns of interest within each image [26]. Saliency maps compute the gradient of the output category (TSC or control) to the original image, that is, they identify the pixels in the original image that, if changed, would modify most the probability of the image belonging to the class [27]. Roughly, class activation maps and saliency maps can be interpreted as “where the CNN is looking at” or, more precisely, which areas in the image are most important for classification into TSC versus control. Map visualizations are heatmaps of the gradients with the “hotter” colors representing the regions of most importance for classification.

Statistical analysis and software

We summarized demographic and basic clinical data with descriptive statistics and calculated basic comparisons with rank sum Wilcoxon test for continuous variables and Fisher’s exact test for categorical variables. For all CNNs, we used Adam [22] as an optimizer with a learning rate of 0.00025, we used binary cross-entropy as loss function, and a batch size of 64 with 100 epochs. We used Python version 3.6 [28] as programming language and within this language, its modules: Pandas [29], NumPy [30] and SciPy.Stats [31], to perform descriptive statistics, Keras [14] and scikit-learn [32] to build and analyze the CNNs, TensorFlow [33] as backend for CNNs, and Keras-vis for visualization of class activation maps and saliency maps [34]. We used Jupyter notebooks to run and save the code and results [35]. We trained and validated our models with anonymized MRI images in a cloud computing system and tested the selected model in our local computer. We used a default threshold of 0.5: if the estimated probability was greater than 0.5, the MRI slice was classified as having tubers, otherwise it was classified as not having tubers.

Demographic and clinical features

The TSC patients and controls included in the study were similar, except TSC patients were slightly younger than controls at the time of imaging (median 9.5 years versus 12.4 years) with no statistically significant difference between the groups (Table 1). The data were divided into 566 images for TSC training (69 patients), 130 images for TSC validation (20 patients), 210 images for TSC testing (25 patients), 561 images for control training (69 patients), 118 images for control validation (20 patients), and 226 images for control testing (25 patients). The training file, merging the TSC and control training images and after data augmentation contained 5,634 images. The validation file, merging the TSC and control validation images contained 248 images. The testing file, merging the TSC and control images contained 436 images.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Demographic features in our population.

https://doi.org/10.1371/journal.pone.0232376.t001

Validation set

The class activation maps in the validation set showed that all the models were classifying images based on the presence of tubers, rather than on other features. The best performing CNN in the validation set was InceptionV3 (loss: 0.5325) followed by ResNet50 (loss: 0.5400), and TSCCNN (loss: 1.0416). Therefore, we used InceptionV3 in the final model. Other performance values are summarized in S1 Table at https://ivansanchezfernandez.github.io/TSC_Supplementary_Table_S1 and also show that InceptionV3 performed better than the other CNN architectures in the validation set.

Evaluation in the test set

InceptionV3 had an accuracy of 0.95 and an area under the receiver operating characteristic curve of 0.99 in the test set with a sensitivity of 0.95, specificity of 0.95, positive predictive value of 0.94, negative predictive value of 0.95, and F1 score (the harmonic mean of positive predictive value and sensitivity) of 0.95 (Table 2).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Performance of InceptionV3 in the test set.

https://doi.org/10.1371/journal.pone.0232376.t002

Model visualization

Although the purpose of this study was not segmentation, which requires a different CNN architecture, the class activation maps and saliency maps showed that the deep learning algorithm was focusing on tubers to classify the images (Figs 1 and 2). All the figures with the Grad-CAM and saliency maps for the 436 testing images can be found at: https://ivansanchezfernandez.github.io/TSC_VisualizationI/#images, https://ivansanchezfernandez.github.io/TSC_VisualizationII/#images, and https://ivansanchezfernandez.github.io/TSC_VisualizationIII/#images. A graph with the colorbar can be found at https://ivansanchezfernandez.github.io/TSC_heatmap_colorbar/.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Correctly classified images.

A. InceptionV3 was able to localize all or most tubers in this image with scattered and sometimes subtle tubers. B. InceptionV3 was able to localize the three relatively well-defined tubers in this image. C. InceptionV3 was able to localize the relatively well-defined tuber in this image. Although the image was classified as having tuber(s), the estimated probability was 0.71, as opposed to >0.99 for A and B. The first column represents the original image, the second column, the map, and the third column the map superimposed on the original image. The first row represents the gradient-weighted class activation map, and the second row represents the saliency map. Both gradient-weighted class activation maps and saliency maps visualizations are based on gradients. The gradient is the partial derivative of the loss function for each pixel in the image of reference (the last convolutional layer for gradient-weighted class activation maps and the original image for saliency maps). Gradient-weighted class activation maps use the gradient of the output category to the last convolutional layer (the last layer with spatial information). Saliency maps use the gradient of the output category to the original image. Both maps methods identify the pixels (in the last convolutional layer for gradient-weighted class activation maps and in the original image for saliency maps) that, if changed, would modify most the probability of the image belonging to the specific class (TSC or control). The resulting visualization is a heat map with values normalized between -1 (purple) and 1 (yellow) with hotter colors representing areas of greater importance for classification (see color bar at https://ivansanchezfernandez.github.io/TSC_heatmap_colorbar/). If you are not familiar with tubers, good examples can be found in Fig 1 in the Peters et al article summarizing neuroimaging in TSC [5]. A version of the images with arrows pointing to the tubers is available as S1 Fig at https://ivansanchezfernandez.github.io/TSC_Supplementary_Figures/.

https://doi.org/10.1371/journal.pone.0232376.g001

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Incorrectly classified images.

We would like to emphasize that incorrectly classified images represented only approximately 5% of the test set, but they sometimes provide insights into the reasons for misclassification. A. InceptionV3 classified this image as having tuber(s) with an estimated probability of 0.82, although it belonged to a control patient. The maps suggest a focus on prominent vascular spaces in the white matter suggestive of radial migration lines. B. InceptionV3 classified this image as having no tuber(s) despite the radiologist-confirmed subtle tuber in the right occipital region. The maps show a focus in the right region, but the model estimated a probability of having tuber(s) of only 4%. C. Although this occurred in a tiny minority of images, this image shows that sometimes the tuber is completely missed and the focus of the maps is not necessarily informative. The estimated probability of having tuber(s) was less than 1%. The first column represents the original image, the second column represents the map, and the third column represents the map superimposed on the original image. The first row represents the gradient-weighted class activation map, and the second row represents the saliency map. Both gradient-weighted class activation maps and saliency maps visualizations are based on gradients. The gradient is the partial derivative of the loss function for each pixel in the image of reference (the last convolutional layer for gradient-weighted class activation maps and the original image for saliency maps). Gradient-weighted class activation maps use the gradient of the output category to the last convolutional layer (the last layer with spatial information). Saliency maps use the gradient of the output category to the original image. Both maps methods identify the pixels (in the last convolutional layer for gradient-weighted class activation maps and in the original image for saliency maps) that, if changed, would modify most the probability of the image belonging to the specific class (TSC or control). The resulting visualization is a heat map with values normalized between -1 (purple) and 1 (yellow) with hotter colors representing areas of greater importance for classification (see color bar at https://ivansanchezfernandez.github.io/TSC_heatmap_colorbar/). If you are not familiar with tubers, good examples can be found in Fig 1 in the Peters et al article summarizing neuroimaging in TSC [5]. A version of the images with arrows pointing to the tubers (except for 2A which had no tubers) is available as as S2 Fig at https://ivansanchezfernandez.github.io/TSC_Supplementary_Figures/.

https://doi.org/10.1371/journal.pone.0232376.g002

Evaluation in clinical cases

To illustrate how readers can apply this deep learning model, we applied it to 6 new patients with TSC and particularly challenging radiological features such as very subtle tubers or age-related incomplete myelination. The deep learning model differentiated MRI slices with tubers from those without tubers with a sensitivity of 0.67, specificity of 0.68, positive predictive value of 0.84, negative predictive value of 0.46, and F1 score of 0.75. See the results and images at: https://ivansanchezfernandez.github.io/TSC_TestCases/.

Interactive model

This deep learning model, the same as the one we used, is packaged in an App so it will produce the same results if exposed to the same images. This App avoids privacy challenges because it has no images on it, just the final model that users can test with their own images. We encourage readers to download and use the App on their own set of TSC test cases. Step-by-step use instructions available for Windows at: https://ivansanchezfernandez.github.io/TSC_TuberFinder_Windows/ and Apple at: https://ivansanchezfernandez.github.io/TSC_TuberFinder_Apple/. This system does not need to train the model, only test it with new images and, therefore, does not need a computer with graphics processing unit (GPU) and takes approximately one minute per image.

Strengths and weaknesses

The present model achieved a classification performance close to that of the gold standard (a clinical neuroradiologist (EY) with formal training in radiology, neuroradiology, and pediatric neuroradiology and part of the epilepsy center at our institution) despite being subject to multiple constraints, which show its robustness: 1) the CNNs were trained on a relatively small dataset, 2) original images were two dimensional .jpg images and their resolution was reduced to 224x224 pixels, a much lower resolution than that used by radiologists to interpret MRIs and of lower image quality, 3) the CNNs were standard architectures developed for multipurpose large object recognition, that is, no specific architectural changes were made to try to fit the CNNs to the task of identifying patterns in MRI images, and 4) the TSC patients were heterogeneous in the number and appearance of tubers in MRI. Despite these challenges, our work shows that CNNs are able to achieve a classification performance close to the gold standard, which may make them able to guide attention to certain MRI slices and areas within each slice where tubers are more likely. CNNs offer the advantage of objectivity: the deep learning model outputs an estimated probability of an image having tubers, and consistency: the deep learning model will always return the same results when presented with the same image. Large repositories of MRI images analyzed with more refined CNNs may further increase sensitivity and specificity to a level where this approach can be used in routine clinical practice. It was not the objective of this study to perform image segmentation. In the future, if a set of MRI images with tuber borders delineated by a radiologist becomes available, other architectures may be used to actually segment tubers. CNN architectures specifically designed for automatic segmentation such as “U”-shaped architectures or fully convolutional neural networks specifically designed for segmentation, may further improve the localization of individual tubers within an MRI slice at the pixel level. Similar to some prior deep learning studies [8, 9, 43, 44], we performed a binary classification task to differentiate between MRI images with tubers and without tubers. The performance of this deep learning model when presented with confounding pathologies (tumors, white matter lesions in multiple sclerosis, etc.) is unknown. In the future, once we gather enough MRI images with other pathologies we aim to develop multiclass classifiers that recognize multiple different pathologies on brain MRI. The performance of this deep learning model (trained only on subjects 5 year old or older) was very good when presented with tubers in the context of a more immature myelination status, as shown by the very challenging test cases (https://ivansanchezfernandez.github.io/TSC_TestCases/).

Deep learning algorithms have been implemented in medicine mostly within the limits of research studies. Their application in routine clinical practice has been challenged by huge computational requirements and privacy concerns. By integrating a trained deep learning algorithm within a standalone application, we demonstrate that application into routine clinical practice and portability of trained CNNs is feasible. Further, as the deep learning algorithm within the application processes images in the local computer, it can be used with patient-protected information as long as the local computer is within a HIPAA-compliant healthcare environment.