Explainable artificial intelligence for differential diagnosis and intervention of prostate carcinoma using multiparametric MRI

Abstract: Multiparametric MRI (mpMRI) combined with convolutional neural network (CNN) is increasingly employed in prostate cancer management. Manual segmentation of the prostate gland and tumors is time-consuming and varies between observers, underscoring the need for automated CNN algorithms. However, the opaque nature of CNNs, often seen as "black boxes," challenges their acceptance in clinical settings. Moreover, integrating mpMRI with clinical data is essential for enhancing diagnostic accuracy and fully utilizing all available information for reliable outcomes. Yet, the optimal method for fusing these data types to achieve the best diagnostic performance remains unclear. This thesis investigates the use of Explainable AI methods to understand the decisionmaking process of CNN for segmenting the prostate gland and prostate tumors. The CNN’s performance was evaluated against two sets of ground truth data: expert derived contours and whole mount histopathology derive contours co-registered with mpMRI. To understand the decision making process of the CNN, 3D Gradient Weighted Class Activation Map, Integrated Gradients, and Layer-wise Relevance Propagation methods were implemented. Further, three data fusion techniques input level, feature level, and decision level were compared for detecting clinically significant prostate lesions. The CNNs were trained with mpMRI images (T2w, ADC, High b-value) and clinical data (PSA, PSAD, prostate gland volume, GTV) to assess influence of fusing these data. The CNN achieved a mean Dice Sorensen Coefficient (DSC) of 0.62 for the prostate gland and 0.32 for prostate tumor segmentation with radiologist-drawn ground truths, and 0.31 against histopathology ground truths. A t-test (p = 0.69) showed no significant performance differences between the two ground truth sets, confirming the robustness of the CNN in identifying relevant segmentation features without systematic bias. The heat maps provided insights into how the CNN distinguishes between tumor and healthy tissues, by localizing and highlighting the most sensitive and relevant pixels for segmentation. For the comparison of data fusion methods, the CNN with decision level fusion outperformed other methods, achieving higher precision, recall, average precision, and F scores. DSC of 0.30, 0.34, 0.36 and 0.26, 0.33, 0.34 showed modest improvements with combined data. However, a t-test (p = 0.26, 0.62, 0.85) showedno significant differences in the performance between the models. Additionally, CNNs trained with different combinations of mpMRI sequences (independently, in pairs and combined) showed that using all three sequences together produced the best results, aligning with clinical PI-RADS protocols

Angaben zum Objekt

• Standort: Deutsche Nationalbibliothek Frankfurt am Main
• Umfang: Online-Ressource
• Sprache: Englisch
• Anmerkungen: Universität Freiburg, Dissertation, 2025

Klassifikation und Themen

• Schlagwort: Kernspintomografie ; Prostatakrebs ; Bildgebendes Verfahren

Beteiligte, Orts- und Zeitangaben

• Ereignis: Veröffentlichung
• (wo): Freiburg
• (wer): Universität
• (wann): 2025
• Urheber: Gunashekar, Deepa Darshini
• Beteiligte Personen und Organisationen: Bock, Michael ; Zamboglou, Constantinos ; Albert-Ludwigs-Universität Freiburg. Medizinische Fakultät

Weitere Informationen

• DOI: 10.6094/UNIFR/262124
• URN: urn:nbn:de:bsz:25-freidok-2621242
• Rechteinformation: Open Access; Der Zugriff auf das Objekt ist unbeschränkt möglich.
• Letzte Aktualisierung: 15.08.2025, 07:38 MESZ