Deep Textural Representations for TNM Staging and Positron Emission Tomography Analysis

Abstract:
Oesophageal cancer is a cancer of unmet need, often presenting late when disease is well established, making clinical management challenging due to complexity and metabolic heterogeneity. This thesis addresses this by harnessing artificial intelligence (AI) to identify and understand Warburg-driven metabolic signatures exhibited in FDG-PET image data.

The first contribution evaluates manifold learning and dimensionality reduction techniques for the interpretation of deep metabolic texture features in PET data. Ten algorithms were assessed using seven quantitative metrics. Results indicate that the choice of evaluation metric substantially affects perceived performance. While non-linear approaches such as UMAP and t-SNE performed well on local clustering metrics, global analysis showed that PCA provided the most distinct and interpretable separation of glycolytic volume features. Image space analysis demonstrated robust and anatomically coherent segmentation of metabolic compartments.

The second contribution compares deep texture features with traditional symbolic radiomics for the identification of glycolytic volumes. Deep features, learned through neural networks, consistently outperformed handcrafted descriptors in both unsupervised clustering and supervised classification. Models based on GNNs achieved the highest accuracy of 97.34% and F1 score of 97.33% in glycolytic volume delineation. These findings indicate that deep texture analysis provides a more accurate and reliable characterisation of glycolytic volumes compared with conventional symbolic radiomics.

The third contribution develops and validates a framework for automated TNM category differentiation in oesophageal cancer, combining supervised deep learning with unsupervised manifold learning. The approach achieved high sensitivity for primary tumours (100%), malignant nodes (81.0%), and metastases (92.0%), with an explainable classification accuracy of 89.5% using unsupervised deep texture arbitration. Expert-in-the-loop validation confirmed relevant lesions missed by standard annotation in 31% of cases.

In summary, this thesis demonstrates that deep learning-based metabolic texture analysis can provide interpretable, quantifiable, and clinically relevant PET biomarkers for oesophageal cancer, supporting improved tumour characterisation and staging.

________________________________________________________________________________