סמינר: Graduate Seminar
Predictive Modeling of Event-Free Survival in Hodgkin Lymphoma Using Radiomics Analysis of Interim PET-CT scans
| Developing robust predictive models for high-dimensional radiomic data remains a significant engineering challenge, particularly in regimes characterized by extreme data scarcity and severe class imbalance. This thesis investigates the efficacy of a multimodal fusion framework designed to integrate heterogeneous data streams, specifically volumetric interim PET/CT radiomics and tabular clinical priors, to predict long-term outcomes.
The proposed framework employs a rigorous signal processing pipeline. First, raw imaging in patients with Hodgkin Lymphoma data undergoes spatial normalization and intensity quantization to ensure feature robustness. To address the variable number of lesions per subject, we implement a volume-weighted embedding strategy, mapping variable-length lesion feature sets into fixed-length patient vectors. We instantiate this framework on the AHOD0831 dataset (N=137) to predict 5-year Event-Free Survival (EFS) in pediatric Hodgkin Lymphoma. This domain serves as a stringent test-bed due to the high dimensionality of the feature space relative to the sample size and a significant class imbalance (19.7% event rate). To isolate the predictive utility of metabolic versus morphological signals, we conducted a systematic ablation study across seven feature configurations. The learning architecture utilized a Nested Cross-Validation (NCV) scheme to prevent data leakage, employing Recursive Feature Elimination (RFE) for dimensionality reduction and optimizing non-linear classifiers (SVM, XGBoost). Experimental results demonstrate that a fusion of CT-derived texture descriptors (specifically NGTDM) and clinical variables, trained on a Support Vector Machine with an RBF kernel, yielded the highest generalization performance. This configuration achieved a balanced accuracy of 0.781 ± 0.096 after Youden index optimization, significantly outperforming metabolic-only baselines. These findings suggest that morphological texture features provide orthogonal information to standard metabolic metrics. Furthermore, this study validates that rigorous feature selection and early fusion strategies can yield stable, explainable predictive models even under the constraints of small, imbalanced medical datasets. M.Sc. student under the supervision of Prof. Moti Frieman.
|