AI Evaluations of Sarcoma From Magnetic Resonance Imaging

Abstract

Background: Sarcomas are a rare and heterogeneous group of malignant tu- mors, making early detection and diagnosis a high priority. Diagnosis tradition- ally relies on expert interpretation of radiological imaging and histopathological biopsies, processes that are time-consuming, subjective, prone to inter-observer variability and are invasive. Methods: This research investigates the potential of artificial intelligence (AI), specifically radiomics and machine learning (ML), to support sarcoma diagnosis, characterization and grading based on MRI scans. Quantitative features were ex- tracted from multiple image transforms, including original, wavelet, Laplacian of Gaussian, square, square root, logarithm, exponential, gradient, and local binary pattern transforms. From these images, first-order statistics and texture descrip- tors (GLCM, GLSZM, GLRLM, NGTDM) were computed. A diverse set of ML models were evaluated, including Random Forest, Logistic Regression, SGDClas- sifier, Ridge Classifier, LightGBM, XGBoost, and CatBoost. Models were trained on three primary tasks: (1) binary classification of healthy vs. sarcoma tissue, (2) sarcoma grading based on the FNCLCC system, and (3) differentiation between bone and soft tissue sarcomas. In addition, a habitat generation framework was introduced, clustering radiomic segments into biologically meaningful subregions to enhance grade and sub-group classification. Model performance was assessed using AUC-ROC, accuracy, and macro- and micro-averaged F1-score, precision, and recall. Results: For binary classification of healthy vs. sarcoma tissue, models per- formed strongly, achieving AUC-ROC scores ranging from 0.829 to 0.999 depend- ing on data preprocessing, with consistently high metric values overall. This demon- strates that ML methods are well suited to distinguishing healthy from patholog- ical tissue. Grade classification achieved AUC-ROC scores ranging from 0.618 to 0.690 using radiomics and between 0.520-0.591 using habitats, with metrics gen- erally above baseline for three-class classification. While performance remains modest, results indicate meaningful correlation between radiomic inputs and FN- CLCC grade. Sub-group classification of bone vs. soft tissue sarcomas achieved AUC-ROC scores of 0.750-0.863 using radiomics and 0.585-0.707 using habitats, indicating that ML methods can effectively aid in sarcoma sub-type differentia- tion. Conclusions: The findings demonstrate that combining multiple image trans- forms with radiomics for diagnostic, characterization and grading classification feasible and effective. Advanced ML models substantially improve performance across diagnostic and prognostic tasks. Furthermore, the habitat generation ap- proach offers additional biological interpretability and can enhance grade and sub- group prediction. These results highlight the promise of AI-driven radiomics pipelines to support clinical decision making in sarcoma diagnosis and hopefully earlier management and extended survival. Keywords: Artificial Intelligence, Machine Learning, Sarcoma, Diagnostics, Sarcoma Grading, Radiomics, Habitats.