Medical imaging

Machine learning (ML) has traditionally been used to make predictive models by training on local data. However, due to concerns regarding privacy, it is not always possible to collect and combine data from different sources. On the other hand, if there are insufficient data available, it might not be possible to construct accurate models to produce meaningful outcomes. This is where Federated Learning comes to the rescue. Federated Learning (FL) represents a sophisticated distributed machine learning strategy that enables multiple devices hosted at different institutions such as hospitals, to collaboratively train a global model while ensuring that their respective data remains securely stored on-premises. It addresses privacy concerns and data protection regulations, because raw data does not need to be shared or centralized
during the training process.
This thesis research studies how two different FL architectures, centralized and decentralized FL, affect medical image classification. To study and validate the findings, skin cancer images dataset is used in a federated learning setting with five sites/clients, and a center for centralized FL. Experimental results show that using both centralized and decentralized (peer to peer) version of FL for classification of skin cancer images outperforms using the traditional ML. In addition, two different FL settings, centralized federated learning (CFL) and decentralized federated learning (DFL), are compared using different data distributions across sites/clients. Our study shows that the best accuracy (95.14%) was achieved with the DFL model when tested on the original dataset (without adding bias to the class distributions). This asserts that class distribution imbalance between sites has a significant impact to the federated learning.

Melanoma, a severe and life-threatening skin cancer, is commonly misdiagnosed
or left undiagnosed. Advances in artificial intelligence, particularly deep learning,
have enabled the design and implementation of intelligent solutions to skin lesion
detection and classification from visible light images, which are capable of performing
early and accurate diagnosis of melanoma and other types of skin diseases. This work
presents solutions to the problems of skin lesion segmentation and classification. The
proposed classification approach leverages convolutional neural networks and transfer
learning. Additionally, the impact of segmentation (i.e., isolating the lesion from the
rest of the image) on the performance of the classifier is investigated, leading to the
conclusion that there is an optimal region between “dermatologist segmented” and
“not segmented” that produces best results, suggesting that the context around a
lesion is helpful as the model is trained and built. Generative adversarial networks,
in the context of extending limited datasets by creating synthetic samples of skin
lesions, are also explored. The robustness and security of skin lesion classifiers using
convolutional neural networks are examined and stress-tested by implementing
adversarial examples.