Graph neural networks (GNNs) have become the standard procedure to deal with graphstructured data and data in non-Euclidean spaces. Since 2017, numerous researchers have been
using GNN models in their experiments. However, despite GNNs’ recent rapid growth, there are not yet many real-world applications that benefit from these models. In this thesis, we use GNNs
as image classifiers. To improve the efficiency and reduce the complexity of the models, we first generate graphs from the images by creating superpixels of the images and use them as our graph nodes instead of individual pixels. Then, we define edges on these nodes based on the distance
of each superpixel to their closest ones. We propose two ensemble frameworks containing a pre-trained ResNet18 and two Graph Neural Network (GNN) models (GAT and GIN). We call these
frameworks GATRes and GATGIN, respectively.
We test these frameworks on two real-world medical applications: Cervical Cancer Screening and Melanoma detection. Cervical cancer is among the top four common cancer in women worldwide.
Cervical cancer can be easily prevented if caught in its pre-cancerous stage. Determining the appropriate treatment method depends on patients’ physiological differences. A treatment that
works effectively for one woman may obscure future cancerous growth in another woman due to differences in the type of their cervix. In this thesis, we experiment with multiple GNNs on this
dataset to distinguish the cervix types and examine if these models enhance detection performance and accuracy.
The other problem we consider is Melanoma skin cancer. Melanoma is the most lethal skin cancer. If melanoma gets diagnosed in the early stages, the patients’ survival rate will increase significantly. This research applies GNN models on the Melanoma dataset to discriminate between melanoma and benign skin lesions.
We show that our models’ sensitivity and accuracy outperform the individual models in our classification tasks. Our GATRes model also outperforms the accuracies achieved by previously
published papers on the Cervical Cancer Screening dataset.