Defense of Jorge Ignacio Facuse Pérez

Abstract

Image-based diagnosis and segmentation with AI methods have considerably improved over recent years thanks to Deep Learning (DL) approaches. However, using DL models on small datasets for populations different from the original training data requires significant amounts of labels, even for the case of fine-tuning a large foundation model previously pre-trained on a larger dataset. Self-supervised learning (SSL) and transfer learning emerge as a solution for this problem. SSL leverages unlabeled data to learn generalizable representations, whereas transfer learning aims at adapting models between different tasks and data distributions and thus facilitates fine-tuning large foundation models on datasets with small labeled data.

This thesis proposes an SSL-based domain adaptation training strategy for training models on small prostate MRI data to perform clinically significant prostate cancer (csPCa) diagnosis and segmentation. This approach involves a self-supervised pre-training stage and a supervised learning training stage on a large MRI dataset, followed by a fine-tuning stage on a smaller target dataset. The publicly available PI-CAI dataset serves as the large dataset, and two smaller out-of-distribution (OOD) prostate cancer MRI datasets (Prostate158 and ChiPCa) were employed to evaluate domain adaptation performance.

We tested several variations of our strategy to find the best training strategies and model architectures for domain adaptation in prostate cancer MRI detection and segmentation. Our results on both datasets validate our proposed training strategy in facilitating the training of DL models on small prostate MRI datasets and reveal that the best strategy among all tested combines UNet as the best model architecture and Denoising Autoencoder (DAE) as the SSL pre-training method, obtaining 0.558 of overall score on the Prostate158 dataset and 0.647 on the ChiPCa dataset.