Radiogenomics combines features from medical images with genomic information to evaluate and classify a pathology. We applied this approach to differentiate between atypical lipomatous tumours (ALTs) and lipomas (which are benign). Imaging was from magnetic resonance (MR) including three types named T1- (makes fat bright), T2- (makes water and fat bright) and fat-suppressed contrast enhanced T1- (makes water bright) weighted sequences. The genomic data consisted of the amplification status of the MDM2 gene; amplification (increase in copy number) of this gene is a hallmark of ALTs. Machine Learning (ML) models (using a LASSO algorithm) were trained on data from persons with ALTs or lipomas (65 and 192, respectively) and known MDM2 amplification status. Testing models (trained on radiogenomics features from all persons) on an additional group of 50 persons indicated discriminant power to detect persons with MDM2 amplification from MR images with better performance (AUC 0.88, 70% sensitivity, 81% specificity, 76% accuracy) than radiology residents (postgraduate trainees) but worse than an attending (more experienced) radiologist [1]. Adding clinical features to the training (age, sex and tumor body region) did not improve performance. These results indicate that the models could help diagnostic work.
In another application of ML to the prediction of clinical outcomes from medical images, we explored the prediction of response to palliative radiation therapy (RT) in painful spinal bone metastases (PSBMs) patients, which works in two thirds of the cases. Planning Computed Tomography (CT) is a specialized CT done for the purpose of strategizing cancer radiation therapy. Planning CT scans of PSBMs were manually segmented (for gross tumour volume, GTV, and clinical target volume, CTV), and used, with semantic and clinical features, to train and compare the performance of random forest (RFC) and support vector machine (SVM) classifiers [2]. We found that, while image-derived features were predictive, semantic features performed similarly, and clinical features were the best predictor. Models combining features did not work better; thus CT scans might not be helpful for the prediction of response to therapy, but we observed that of the radiomics features, CTV segmentations, which include vertebra compartments that are at risk of microscopic infiltration, predicted better than GTV segmentations.
Application of ML for the prediction of the side effects of breast cancer RT (skin inflammation and subsequent moist epitheliolisis and edema) was used in planning CT from 252 patients [3]. ML models were trained on features from total breast volume (TBV; encompasses all mammary tissues) or from glandular tissue (GT; one of several mammary tissues). Prediction of RT side effects was better for models trained with TBV image features, and the volume of the TBV correlated with better predictive performance. Exclusion of features correlated with volume reduced predictive power only slightly. In conclusion, the features derived from planning CT were predictive, but less than TBV volume.
References
[1] Foreman, S.C., O. Llorian-Salvador, D.E. David, V.K.N. Rösner, J.F. Rischewski, G.C. Feuerriegel, D.W. Kramp, I. Luiken, A.K. Lohse, J. Kiefer, C. Mogler, C. Knebel, M. Jung, M.A. Andrade-Navarro, B. Rost, S.E. Combs, M.R. Makowski, K. Woertler, J.C. Peeken, A.S. Gersing. 2023. Development and evaluation of MR-based radiogenomic models to differentiate atypical lipomatous tumors from lipomas. Cancers. 15, 2150.
[2] Llorián-Salvador, O., J. Akhgar, S. Pigorsch, K. Borm, S. Münch, D. Bernhardt, B. Rost, M.A. Andrade-Navarro, S.E. Combs, J.C. Peeken. 2023. The importance of planning CT-based imaging features for machine learning-based prediction of pain response. Sci. Rep. 13, 17427.
[3] Llorián-Salvador, O., N. Windeler, N. Martin, L. Etzel, M.A. Andrade-Navarro, D. Bernhardt, B. Rost, K.J. Borm, S.E. Combs, M.N. Duma and J.C. Peeken. 2024. CT-based radiomics for predicting breast cancer radiotherapy side effects. Sci. Rep. 14, 20051.