| dc.description.abstract | The detection of a lesion in the prostate can be a challenging task and is crucial for the early diagnosis of Prostate Cancer (PCa). Magnetic Resonance Imaging (MRI) examination provides a comfortable and precise solution to detect prostate lesions. The ability of humans to detect lesions from the prostate MRI by learning from the appearance of healthy prostate structures might help deep learning (DL) architectures achieve the
human level’s detection ability.
To this end, this thesis proposes an effective method to detect lesions in the patients by learning the distribution of healthy prostate images using auto-encoder-based methods in an unsupervised framework. The thesis methodology involves two main steps: training of DL models, and binary classification of the images as well as detection of lesions in unhealthy images. This work makes use of two DL architectures for the task: Variational Autoencoder (VAE) and Autoencoders (AE), which are then compared in terms of lesion detection and classification ability. The binary classification is based on pixel-wise reconstruction error. The thesis uses the T2w and Apparent Diffusion Coefficient (ADC) MRIs of the prostate, from PROSTATEx Challenge data. The thesis explores the effect of data imbalance in the final results by using two different configurations of test data, balanced and imbalanced data, for both modalities.
The final results indicate that VAE performs significantly better than AE in terms of ROC-AUC, and both models perform notably better for ADC images than T2w images. | |
