Unsupervised anomaly detection for tumor delineation in a preclinical model of glioblastoma using CEST MRI

Glioblastoma is characterized by heterogeneous tumor characteristics and infiltrative tumor boundaries, making accurate delineation difficult with extensive manual annotations. Chemical exchange saturation transfer (CEST) is a non-invasive MRI technique used for in vivo assessment of metabolic and macromolecular information through a Z-spectrum. CEST may provide insight into metabolic changes present in early-stage disease that are not visible in routine clinical imaging, thereby improving tumor delineation. In this work, we use an unsupervised anomaly detection (UAD) strategy to learn the distribution of features present in Z-spectra of healthy tissue and capture their deviations in pathology, foregoing the need for extensive labels. The approach leverages the metabolic information provided by CEST to improve the detection and delineation of glioblastoma and inform further treatment planning.

A 1D convolutional autoencoder (CAE) was implemented to reconstruct Z-spectra from individual tissue voxels. The network was trained on Z-spectra acquired at 9.4T from healthy Sprague-Dawley rats and tested on data acquired from F98 glioma-bearing rats post Gd-administration. For baseline comparisons, Isolation Forest and Local Outlier Factor, which have shown success in anomaly detection, were implemented. For the CAE, our anomaly score was determined to be the mean squared reconstruction error. To facilitate clinical translation and evaluate the robustness of our model for under sampled Z-spectra, acceleration factors of 2x and 7x were performed with two sampling schemes: uniformly skipping frequency offsets and selecting offsets based on feature importance identified by Shapley value analysis and Integrated Gradients (IG). Binarization was performed by determining an optimal anomaly threshold, followed by comparison to ground truth tumor masks. Metrics related to model performance were assessed for baseline anomaly detectors on the fully sampled dataset and for the CAE on fully and under sampled datasets.

The best baseline anomaly detector was Isolation Forest, with an ROC-AUC of 0.967 and an F1-score of 0.584. Our method, the CAE, accurately reconstructed Z-spectral features, achieving Dice scores of up to 0.72 and outperforming the baseline model with an ROC-AUC of 0.968 and F1-score of 0.642. This model performance remained robust across sampling schemes and acceleration factors, with ROC-AUCs of ~0.96 and similar Dice scores (up to 0.7). Feature importance analysis indicated that offsets in the range of ±3.0 to 5.0ppm contributed most to the anomaly score.

This study successfully demonstrated a UAD pipeline utilizing the Z-spectrum from CEST MRI for metabolically informed tumor delineation. The framework captures biochemical deviations that may precede or extend beyond morphologic abnormalities, enabling sensitive detection of tumor regions and intra-tumoral heterogeneity that previous methods may fail to capture. The offsets from the feature analysis indicated a strong contribution from the magnetization transfer (MT) pool to the spectral deviations captured by the model, with additional contributions from relayed nuclear Overhauser effect (rNOE) and amide proton transfer (APT). Model robustness with under sampling further highlights the pipeline's potential in accelerated acquisitions, thus improving clinical practicality. While there is a need for validation on larger cohorts and clinical datasets, the current results demonstrate that this label-free, Z-spectral anomaly mapping can serve as an interpretable and scalable tool for monitoring tumor heterogeneity and progression, with potential applicability to other diffuse or metabolically subtle pathologies.