Skull Base Chordoma: Automated Detection and Segmentation Using Radiomics

Introduction: Skull base chordomas are rare, locally destructive, and recurrent pathologies, originating from remnant of notochord in the sphenooccipital synchondrosis. Skull base chordoma share similar location, radiologic features, growth pattern, and clinical presentation with chondrosarcoma. The radiological distinction between chordoma and chondrosarcoma/or postoperative changes is often challenging.

In our study, we aimed to create a radiomics model solving these problems:

• Distinction of the pathology (clival chordoma/chondrosarcoma).

• Distinction of the chordoma recurrence/residue from postoperative/postradiotherapy changes.

• Automatic detection and segmentation of chordoma.

• Dichotomization of the chordoma into low-risk/high-risk tumor residues.

Methods: This study included 36 patients operated for clival chordoma, 9 chondrosarcomas. Anonymized, pre-processed pre-/postoperative MRIs were collected for each patient: CET1-weighted MRI, T2-weighted MRI, FLAIR.

Due to the very low number of patients in the study, we could not apply solely neural network (NN), instead we performed combined radiomics, including feature extraction, and binary classification models (LASSO, boosting methods, and sequential NN (multilayered perceptron, MPL). The MPL was feed forward SNN including three layers of neurons with RELU function. More than 4,000 combined features were calculated for each patient.

Semiautomated segmentation was initiated with 5 × 5 voxel manual “seed,” expanded using our active “contour” algorithm, and filtered by feature-map (highest Shapley score). The “seed” in the fully automated detection of chordoma was intercept within the predefined midline FOV by feature-based map. Dice coefficient score was calculated for predicted and ground truth segmentation. All processes were provided by our algorithm written in Python 3.8.8 (SimpleITL library, Python, Keras).

Results:

1. Distinction of histology class:

The distinction of the skull base chordoma from chondrosarcoma from the preoperative imaging of three sequences was performed on the augmented cohort including 45 patients.

The best performing radiomics model was XGBoost model based on extraction of combined features, with MSE: 0.175, AUC: 0.77, accuracy = 0.81(95% CI: 0.74–0.94). Three-layered MLP achieved the accuracy of 0.875, MSE 0.195, but the unbalanced complexity of the model is visible from training and validation loss.

2. Distinction of real chordoma residue from postoperative changes:

Model was trained on 21 chordoma residues versus 30 randomly created ROI “fake residues” in known postoperative cases. The best performing XGBoost model achieved performance: MSE: 0.14, AUC: AUC = 0.87, accuracy = 0.8(95% CI: 0.66–0.95).

3. Automated segmentation part:

The automated segmentation model based on “manual seed” achieved mean DICE coefficients 0.92, 0.94, and 0.88 for tumor based on T2-weighted MRI, CET1-weighted MRI, and FLAIR sequences, respectively (Fig. 1).

Automated segmentation part with automatic detection of seeding voxels: The mean DICE scores for automatic segmentation were 0.85, 0.89, and 0.82 for T2-weighted MRI, CET1-weighted MRI, and FLAIR sequences, respectively.

4. Dichotomization into high-risk/low-risk chordoma from postoperative MRI of residue:

The best performing radiomics model was XGBoost of all three sequences, with performance metrics: MSE: 0.12, AUC = 0.96, accuracy 0.8 (95% CI: 0.59–0.94; Fig. 2).

Conclusion: Despite limitations such as small size of dataset and heterogeneous origin of imaging, our MRI-based machine learning model distinguishes with a good accuracy the chordoma from chondrosarcoma, or postoperative changes. The model of automated segmentation based on the intercept of feature based-matrix and active contour algorithm had a surprisingly high accuracy in comparison to ground-truth segmentation.

Publikationsverlauf

Artikel online veröffentlicht:
07. Februar 2025