Deep learning-based prostate cancer detection in magnetic resonance imaging

Abstract

This thesis explores the application of deep learning (DL) models to improve the detection and diagnosis of clinically significant prostate cancer (csPCa) in T2-weighted magnetic resonance imaging (MRI) scans. The primary focus is on developing and comparing the ability of convolutional neural network (CNN) and vision transformer (ViT) models to identify and classify csPCa accurately.

The DL models were trained and evaluated on the public Prostate Imaging: Cancer AI (PI-CAI) dataset, which comprises T2WI scans along with clinical annotations. A private dataset from Stavanger University Hospital (SUS) was used for external validation of the models.

Results indicate that the DL models show potential in detecting csPCa. However, there are problems with low detection rates and numerous false positives. A comparison between the CNN and ViT models reveals no significant performance differences. Both models yielded an AUC of 0.53 on the patient level of the PI-CAI dataset, and an AUC of 0.56 and 0.55, respectively, on the SUS dataset. An explainability analysis shows that the models to some degree can identify relevant image regions.

To increase the performance of the DL models, several improvements are suggested. These include the use of multi-modal data to enhance detection capabilities, the application of location-based attention mechanisms to focus the models on relevant image regions, and the development of models that can segment and locate lesions.

This thesis explores the application of deep learning (DL) models to improve the detection and diagnosis of clinically significant prostate cancer (csPCa) in T2-weighted magnetic resonance imaging (MRI) scans. The primary focus is on developing and comparing the ability of convolutional neural network (CNN) and vision transformer (ViT) models to identify and classify csPCa accurately.

The DL models were trained and evaluated on the public Prostate Imaging: Cancer AI (PI-CAI) dataset, which comprises T2WI scans along with clinical annotations. A private dataset from Stavanger University Hospital (SUS) was used for external validation of the models.

Results indicate that the DL models show potential in detecting csPCa. However, there are problems with low detection rates and numerous false positives. A comparison between the CNN and ViT models reveals no significant performance differences. Both models yielded an AUC of 0.53 on the patient level of the PI-CAI dataset, and an AUC of 0.56 and 0.55, respectively, on the SUS dataset. An explainability analysis shows that the models to some degree can identify relevant image regions.

To increase the performance of the DL models, several improvements are suggested. These include the use of multi-modal data to enhance detection capabilities, the application of location-based attention mechanisms to focus the models on relevant image regions, and the development of models that can segment and locate lesions.