Use Case: Brain Tumor Detection

With the growing use of machine learning (ML) in healthcare, we set out to create an image classification model that could potentially help doctors identify brain tumors.

The model we came up with is trained to take a brain scan image as input and classify whether or not a brain tumor is present in the image.

Dataset

We grabbed the training images from this Kaggle project and pre-processed each of them into a resolution of 232x300 pixels. Below are a couple of examples of these images:

Figure 1: Example images from the training dataset. The image on the left clearly shows a brain tumor.

Creating the Model

To use the images in PerceptiLabs, we first created a .csv file that maps the images to the labels: yes (i.e., contains a brain tumor) and no (does not contain a brain tumor). The following shows an example of what this .csv file contains:

images

labels

yes/Y38.jpg

yes

no/no 3.jpg

no

This .csv file and the brain scan images are available on GitHub.

We then loaded the .csv file into the Data Wizard, setting the first column to images and the second column to categorical. We used the default data partitioning of 70% training, 20% validation, and 10% test, with randomization enabled, and left the default training values provided by the Data Wizard. The goal at this point was to generate a model that we could then modify.

Model Summary

After the model was generated, we went to work replacing the Components to end up with the following model architecture:

---------------------------------------------------------------------------------------------------------------------

Component 1: VGG16

---------------------------------------------------------------------------------------------------------------------

Component 2: Dense Activation=ReLU, Neurons=128

---------------------------------------------------------------------------------------------------------------------

Component 3: Dense Activation=Softmax, Neurons=2

---------------------------------------------------------------------------------------------------------------------- The following shows what this model looks like in PerceptiLabs:

Figure 2: The model architecture in PerceptiLabs.

When then trained the model with the following settings:

  • Epochs: 10

  • Batch size: 4

  • Loss: Cross-Entropy

  • Learning Rate: 0.001

  • Optimizer: ADAM

With a training time of less than 10 minutes, we achieved a training accuracy of 96.05% and a validation accuracy of 90%. In the following screenshot from PerceptiLabs you can see how the training accuracy ramped up over the first three epochs and then stabilized, while validation accuracy dipped around the third epoch before rising again and stabilizing:

Figure 3: Plot in PerceptiLabs comparing training and validation accuracy across 10 epochs.

Vertical Applications

In the realm of diagnosing brain tumors, a model like this could be used to help automate the process of examining brain scans and to notify doctors as to which cases may require a closer look.

Such a project could also be used by medical students or practitioners looking to build next-generation ML-based medical technology. The model itself could also be used as the basis for transfer learning to create additional models for detecting other types of tumors.

Live Coding Demonstration

In the following Live Code Demonstration, originally streamed May 27, 2021, we demonstrate how to set up this model and experiment with different neural networks to achieve different results: