Machine Learning based lesion detection using 18F-FDG PET/MR

Abstract

This thesis presents a pipeline for automated lesion detection in 18F-Fluorodeoxyglucose (18F-FDG) PET/MR images, aimed at addressing diagnostic challenges in chronic pain patients. Due to the limited availability of annotated PET/MR datasets, especially for subtle or rare lesion types, we propose a synthetic lesion data generation method based on reduced-dimensional representations of lesion features derived through Principal Component Analysis (PCA). We generate synthetic lesions by extracting an independent feature basis via PCA from real lesions and sampling from this space using Gaussian Mixture Models (GMMs). These anatomically plausible lesions are inserted into healthy PET images at clinically relevant locations delineated by TotalSegmentator. By varying lesion count and characteristics, we constructed an extensive annotated dataset for training a nnU-NetV2 model. Our experiments show that models trained with synthetic lesions outperform those trained exclusively on real lesions in detecting unseen real cases, demonstrating the power of synthetic augmentation when data is scarcely available. A systematic feature analysis reveals that beyond intensity-based characteristics such as peak SUV and spatial decay, lesion convexity is a crucial factor influencing detectability. In addition, we explore the use of deformation fields to introduce anatomical variations in datasets. By registering patient pairs using rigid, affine and B-spline transformations by their segmentation masks, we enable the generation of anatomically diverse training data without manual annotation. This approach is especially useful when applying it to public datasets, as it allows for the introduction of new anatomies in the training set enhancing the robustness of trained deep learning models. We have shown the potential of this method to prevent false positives in lesion detection. This work lays the foundation for robust AI-driven lesion detection in PET/MR imaging, with potential extensions toward anatomically complex structures like nerve roots and facet joints. The ability of the pipeline to generate delineated lesion masks also opens potential for downstream applications such as lesion quantification, radiomic analysis and treatment planning support.