Abstract—We attempt to classify cancer prevalence by region using deep learning. We used the existing pre-trained VGGNet object recognition model as a base, and applied transfer learning to classify regions by their cancer prevalence. We propose a model that combines the pre-trained VGGNet model for extracting higher level features, and two additional linear layers to extract a relationship between those features and the cancer rates. We show that this model is able to successfully classify areas by cancer prevalence given the satellite images of that area as input.

In the last several years, as the amount of satellite imagery available to the public has increased and deep neural nets have become much more powerful, the potential of training models on these images has become apparent. Satellite images contain a lot of information, but it is often too time consuming for humans to go through and manually extract information from them. Recently, many works have been published with the purpose of extracting useful features from satellite images. In particular, we are interested in the potential of using satellite images as an analogue for the built environment. The built en- vironment is believed to influence a variety of health outcomes. Training a network on these satellite images could provide valuable information about the built environment, which in turn could be correlated with specific health outcomes.

One similar paper published by Maharana and Nsoesie used transfer learning on satellite images with VGGNet to extract features from the images. Essentially, they used a pre- trained object recognition model which had been trained on other datasets, and ran the satellite images through the model with the idea that the resulting output would contain useful information about the dataset. They then trained a linear layer to correlate these features with obesity prevalence. They found that those features explained 72% to 90% of the variation in obesity rates between cities [1]. We plan to use a similar approach, except with cancer rates as a health outcome instead of obesity rates. We will use satellite images from different regions of the United States and extract features from these images using a neural network. We plan to use this neural network to predict cancer rates on average based on the images of their neighborhood, since there exist geographic disparities in prevalence. We will get satellite images and already categorized regions of high or low prevalence of cancer prevalence to train and validate the results of our neural network. Our final goal is to produce an accurate deep learning model to predict the degree of cancer prevalence in a region based solely on the features extracted from satellite images.

We utilized two different data sets in order to train and run our model. Firstly, we needed information about the prevalence of cancer in different locations throughout the United States. The 500 Cities database from the CDC provides a prevalence score for the disease in cities all over the country. Each location is associated with a geolocation, meaning longitude and latitude, a state, a city, and prevalence score for cancer. There were multiple datapoints within a single city. The second part of data collection was finding satelling imagery for all of the locations that the 500 Cities database gave us information about. In order to collect these images, we used the Google Static Maps API to download a satellite image at each of the geolocations listed by the 500 Cities database that had cancer prevalence scores. Using the Google Static Maps API we were able to specify zoom level, size of image, and and exact coordinates.

The 500 Cities database has a heat map that shows preva- lence of cancer in each city. This helps to visualize the prevalence and how it relates to the satellite images.

Fig. 1. Heatmap for Prevalence in Baltimore. Shows the prevalence in each section of Baltimore.

Fig. 2. Satellite Image of Baltimore. Shows the satellite image for Baltimore as downloaded from Google Static Maps API.

With regards to the 500 Cities data that we collected, we had to perform a few pre-processing steps to make the data usable. We removed all data points that did not have a cancer prevalence score. The CDC did not publish prevalence scores for places where the population was very small, so as not to skew the data with small sample sizes. In addition, instead of using the prevalence scores, as a continuous variable, we made a new variable which was a prevalence category, which has six, equally sized, ranges from the lowest prevalence score to the highest prevalence scores. We thought it would be too difficult for the model to predict prevalence scores to the nearest tenth, as they were given in the 500 Cities database, so to make the task simpler, we categorized them into six buckets. For the satellite images, we sized them to be 224 by 224 since that was the requirement for the pre-trained VGG network that we used. In addition, we set the zoom level such that features such as buildings, roads, bodies of water, fields and such were visible by eye. This was to ensure that the deep network would be able to extract meaningful features from the satellite images.

For our deep network we used the VGG-16 model that has been pretrained on Imagenet. The exact architecture is shown below. It a series of convolutional and max-pooling layers, followed by several linear layers to map all the features to an output. We used transfer learning in order to use the VGG network to extract features from the satellite images, and then train one or two new linear layers at the end to map these features to a prediction of cancer prevalence. The original architecture of the VGG model. We experimented using training one new layer at the end, and then training two linear layers to see which results in the best accuracy. In addition we tested two different pre-trained VGG networks, one with batch normalization, and one without to see which gave the best performance.

D. Loss Function Since we formulated the problem as a regression, we experimented with two different loss functions to calculate a distance between the output of the model and the correct label. First we tried an L1 loss function, then we tried L2, or mean squared error loss function. These two loss functions are shown below. For our model, the loss, during training, was only back-propagated through the last layer or two after the original VGG pretrained model.

We first split the data into 80% training and 20% testing. From the training data, we extracted the geolocation for each data point, and used the Google API get the satellite image of that region. All of these images were downloaded. Then during the training process, each satellite image was passed through our model, and a prevalence score (1-6) was generated as output. This was compared against the known prevalence score, which was the label, and a loss was computed and step was taken by the optimizer. We used the Adam optimizer for training. We had around 30,000 locations, and thus that many images as well for our dataset.

After the training the model in the way described, the results are shown in figure 4. First we compared training one linear layer at the end versus training two layers using trans- fer learning. Using two linear layers improved performance considerably; the validation loss reached a smaller value than for one linear layer. Adding more layers improves the fit of our model to the data since there are more parameters that can be adjusted, and in general, making a model deeper makes it better at fitting the given data. The accuracy goes up significantly when we added and trained two layers versus just adding and training one. Next we tests running 10 epochs versus running 20 epochs. As shown, the validation loss goes down for almost all 20 epochs, and then seems to start stabilizing at the end of the experiment. Since our validation loss goes down during the entire experiment, we can conclude that we are not overfitting the training data, and our model is generalizing well to the validation data set, because the validation loss, or training loss, do not start increasing at any point. The accuracy increased when we increased the epoch count from 10 to 20 as well, as expected. We tests two different loss functions, L1 and L2, and as shown in the figures the L2 loss function better, at least with regards to validation and training loss. The only issue we thought of is that it could be possible that the L2 loss, or mean squared error, results in the algorithm just predicting the mean, just to reduce the mean squared error, so we tested the accuracy to see if this was the case. We know this is not the case because the accuracy of our validation set was better when we used

Fig. 3. Network Architecture.This architecture is the VGG-16 model that has been pretrained on Imagenet. It outputs a vector of features that is 1x1x1000. We use one or two linear layers to reduce this to 1x1x1.

L2 loss versus when we used L1 loss. Finally we tested the VGG pretrained network with batch normalization versus without. In this case, batch normalization hurt the performance of the model.

Validation Accuracy
Model Layer Loss Epochs Accuracy
VGG 	1     L1 10     20.26%
VGG-BN 1      L1 10     22.67%
VGG 	1     L1 20     37.34%
VGG 	1     L2 10     45.70%
VGG 	2     L1 20     55.97%
VGG 	2     L2 20     62.39%

Validation Accuracy for Different Models Tested

Our aim was to create a deep network that can accurately predict cancer prevalence based on features of satellite images. We wanted to develop a model that can map geographical features from those images to cancer prevalence. We used the VGG-16 model architecture for the deep network that has been pretrained on Imagenet, and we trained the last few layers of the model to be able to map the features that the net extracted to a prevalence prediction. From our results, it is clear that the model was able to learn a relation between features of satellite images and prevalence. Our the validation loss on all of our experiments did go down, so the training process allowed the model to learn these features. In addition, the accuracy from the bottom three networks in Table 1 show that the model was successful in predicting cancer prevalence from satellite image features.

###VI. FUTURE WORK In the future, we would like to see what features of the satellite images correlate to health outcomes. Our results could be improved additionally by performing further fine tuning on the VGG model. For example, we only re-trained one or two additional layers. It's likely that re- training more layers of the model would improve the accuracy of the model. We also may want to experiment with different satellite resolutions to see what yields the best results. For example, a more zoomed in picture may be better since it would allow the model to identify more detailed features such as roads and vehicles. Additionally, the VGGNet is used for individual object recognition within standard photos. While we were still able to achieve good results with satellite imagery, using a model which was trained to identify features in satellite imagery in the first place instead of normal photos might yield better results. In the paper published by Albert, Kaur, and Gonzlez in 2017, they used a dataset of satellite images of various cities in Europe to train a model used to identify features in satellite images [2]. Using a model like this, or retraining the last few layers of the VGG model on a similar dataset might lead to a better result. Finally, other health outcomes in the 500 cities dataset could be explored using the same approach. Cancer rates may not correlate extremely well with the built environment no matter how fine tuned or complex the model is. Therefore, examining other health outcomes may lead to better and more meaningful results.

Fig. 4. Test Loss Plots. This figure shows the test loss for each experiment that we tried.