Generative Deep Learning for Cancer Image analysis
[eng] Accurate and timely detection of cancerous lesions in medical imaging is essential for effective treatment. However, the diagnosis remains challenging due to, among others, tumor heterogeneity, imaging constraints, and observer variability. Deep learning architectures, such as convolutional an...
| Autor: | |
|---|---|
| Tipo de recurso: | tesis doctoral |
| Estado: | Versión publicada |
| Fecha de publicación: | 2025 |
| País: | España |
| Institución: | Universidad de Barcelona |
| Repositorio: | Dipòsit Digital de la UB |
| OAI Identifier: | oai:diposit.ub.edu:2445/227683 |
| Acceso en línea: | https://hdl.handle.net/2445/227683 http://hdl.handle.net/10803/696841 |
| Access Level: | acceso abierto |
| Palabra clave: | Intel·ligència artificial Processament d'imatges Imatges mèdiques Processament de dades Aprenentatge profund Artificial intelligence Image processing Imaging systems in medicine Data processing Deep learning (Machine learning) |
| Sumario: | [eng] Accurate and timely detection of cancerous lesions in medical imaging is essential for effective treatment. However, the diagnosis remains challenging due to, among others, tumor heterogeneity, imaging constraints, and observer variability. Deep learning architectures, such as convolutional and transformer-based networks, have been showing promise in improving cancer image analysis by learning complex patterns within features from raw imaging data, allowing for earlier, more precise detection and consistent, data-driven decision-making. Despite its potential, clinical adoption of deep learning is restricted by the need for large, annotated training datasets, which are scarce due to privacy and labeling cost constraints, as well as due to its variability in performance when applied in settings of domain shift, imaging artifacts, variation in imaging protocols, and patient populations. This thesis identifies and addresses the challenges in deep learning for cancer imaging through five core publications that propose novel frameworks, methods, and solutions. First, a large-scale survey of generative models in cancer imaging is conducted leading to the identification of key challenges and problems in the field alongside ideation of potential solutions. Based on these findings, the SynTRUST meta-analysis framework is derived to assess the trustworthiness and maturity level of cancer image synthesis studies and solutions. Second, conditional generative adversarial networks (GANs) are applied to the challenges of scarcity of cancer images and tumor annotations, by simulating dynamic contrast-enhanced breast magnetic resonance imaging (DCEMRI) sequences. Without relying on physical contrast agents or DCE-MRI data from real patients, this approach enables unsupervised tumor detection, localization, and characterization, along with providing synthetic training data for increasing the robustness of downstream task models such as tumor segmentation models. Third, a multi-conditional latent diffusion models is developed to translate non-contrast enhanced MRI images into variable time-dependent synthetic DCE-MRI images localizing tumors and predicting their contrast enhancement kinetic patterns. Addressing the need for respective quantitative evaluation metrics, the Fréchet Radiomics Distance (FRD) is proposed to measure (synthetic) image quality based on biomarker variability. Fourth, a mass malignancy-conditioned generative adversarial network (MCGAN) is proposed to generated synthetic data as privacy-preservation mechanism for training deep learning models without individual patient data. Via comparison and combination with differential privacy, the synthetic mammography data is shown to improve the performance in multiple privacy-preserving cancer classification scenarios. Fifth, the medigan library is introduced as sharing platform for pretrained generative models that enable researchers to generate high-quality synthetic data across diverse imaging modalities without requiring direct access to sensitive patient data. Additionally, medigan’s generative models are comprehensive analyzed based on the Fréchet Inception Distance (FID) using both radiology domain-specific and standard domain-invariant feature extractors. In conclusion, this thesis highlights the potential of generative deep learning to address the key challenges in cancer imaging, presenting novel methods in contrast enhancement simulation, tumor localization, privacy-preserving cancer classification, and tools for sharing and assessing generative models, paving the way for these models towards integration into clinical practice to the end of advancing healthcare for both individual patients and society at large. |
|---|