Metastases Detection in Lymph Nodes using Transfer Learning

Metastases detection using Artificial Intelligence can outperform the human but requires extensive training time and resources. In this thesis it is investigated to what extent we can exploit state-of-the-art AI models to effectively learn metastases detection in different lymph node regions. A deep convolutional neural network is trained on the CAMELYON16 challenge as a source model. This network is finetuned to detect metastases in colon or head-neck regions using minimal resources and trying to prevent from catastrophic forgetting. We demonstrate that transfer learning can learn a related task with a small amount of training data and that it can even be used to increase performance on the source task. Parameter values as the learning rate and dataset size are highly task-specific. Elastic Weight Consolidation can be used to prevent from catastrophic forgetting at a minimal cost. All this enables learning a task with less training time and resources.