| Name: | Description: | Size: | Format: | |
|---|---|---|---|---|
| 9.19 MB | Adobe PDF |
Authors
Advisor(s)
Abstract(s)
Neoadjuvant therapy (NAT) has significantly improved outcomes in PDAC, yet patient responses vary considerably, highlighting the need for reliable predictive biomarkers. We developed an interpretable framework using 3D convolutional autoencoders (CAEs) to learn pre-NAT latent features for survival modeling and difference-map features (pre-post) for response modeling in locally advanced pancreatic cancer (LAPC), with clustering and explainability to relate imaging phenotypes to clinical/radiomics data. CTs were preprocessed, a CAE was pre-trained on the PANORAMA dataset (n=676) and finetuned on the LAPC cohort (LAPC, n=127). Survival pipeline: three CAE configurations were trained from pre-NAT scans: (1) scans only; (2) scans+clinical; (3) scans+clinical+Pre-NAT radiomics (PCA). Response pipeline: three CAE configurations from pre- and post-NAT scans were trained: (1) scans only; (2) scans+clinical; (3) scans+clinical+delta radiomics (PCA). Performance was estimated with stratified 5-fold CV. We clustered embeddings from the best CAE models using K-means, tested associations with clinical variables, PCA radiomics, and CA 19-9, and explained cluster drivers with UMAP plus a random-forest surrogate and SHAP. In both pipelines, adding clinical+radiomics during training/fine-tuning improved the latent space geometry (higher Silhouette score, lower Davies-Bouldin). Post-training, K-means on embeddings showed: (i) response pipeline – moderate separation but no association with CA 19-9 change or survival; clusters and SHAP aligned with delta radiomics PCs, not routine clinical variables; (ii) survival pipeline – coherent pre-NAT latent geometry with clusters that were non-prognostic; clusters and SHAP aligned with pre-NAT radiomics PCs, again weakly with clinical variables. The proposed CAE framework yields interpretable, spatially localized representations that capture treatment effect and partly overlap with handcrafted radiomics; however, unsupervised clusters from these embeddings did not translate into prognostic groups in this cohort. Next step: outcomeguided multimodal learning (shared latent space for survival and response), CA 19-9 trajectory modeling, and aligned pre/post CTs with tumor-focused analysis to convert these interpretable signals into actionable biomarkers.
Description
Tese de mestrado, Engenharia Física, 2025, Universidade de Lisboa, Faculdade de Ciências
Keywords
Locally Advanced Pancreatic Cancer Unsupervised Clustering Convolutional Autoencoder Delta radiomics Interpretable AI