machine-learning-algorithms-for-early-detection-bone-metastases-an

Machine Learning Algorithms for Early Detection of Bone Metastases in an Experimental Rat Model

Machine learning (ML) algorithms permit the integration of different features into a model to perform classification or regression tasks with an accuracy ...

Friedrich-Alexander Universität Erlangen-Nürnberg

This study aimed to determine the potential degree of radiation dose reduction achievable using Advanced Modeled Iterative Reconstruction (ADMIRE) in abdominal computed tomography (CT) while maintaining image quality. Moreover, this study compared differences in image noise reduction of this iterative algorithm with radiation dose reduction.

Advanced Modeled Iterative Reconstruction (ADMIRE) Facilitates Radiation Dose Reduction in Abdominal CT.

Macrometastases in bone are preceded by bone marrow invasion of disseminated tumor cells. This study combined functional imaging parameters from FDG-PET/CT and MRI in a rat model of breast cancer bone metastases to a Model-averaged Neural Network (avNNet) for the detection of early metastatic disease and prediction of future macrometastases. Metastases were induced in 28 rats by injecting MDA-MB-231 breast cancer cells into the right superficial epigastric artery, resulting in the growth of osseous metastases in the right hind leg of the animals. All animals received FDG-PET/CT and MRI at days 0, 10, 20 and 30 after tumor cell injection. In total, 18/28 rats presented with metastases at days 20 or 30 (64.3%). None of the animals featured morphologic bone lesions during imaging at day 10, and the imaging parameters acquired at day 10 did not differ significantly between animals with metastases at or after day 20 and those without (all p > 0.3). The avNNet trained with the imaging parameters acquired at day 10, however, achieved an accuracy of 85.7% (95% CI 67.3-96.0%) in predicting future macrometastatic disease (ROC 0.90; 95% CI 0.76-1.00), and significantly outperformed the predictive capacities of all single parameters (all p ≤ 0.02). The integration of functional FDG-PET/CT and MRI parameters into an avNNet can thus be used to predict macrometastatic disease with high accuracy, and their combination might serve as a surrogate marker for bone marrow invasion as an early metastatic process that is commonly missed during conventional staging examinations.

Prediction of early metastatic disease in experimental breast cancer bone metastasis by combining PET/CT and MRI parameters to a Model-Averaged Neural Network.

We investigated whether the integration of machine learning (ML) into MRI interpretation can provide accurate decision rules for the management of suspicious breast masses. A total of 173 consecutive patients with suspicious breast masses upon complementary assessment (BI-RADS IV/V: n = 100/76) received standardized breast MRI prior to histological verification. MRI findings were independently assessed by two observers (R1/R2: 5 years of experience/no experience in breast MRI) using six (semi-)quantitative imaging parameters. Interobserver variability was studied by ICC (intraclass correlation coefficient). A polynomial kernel function support vector machine was trained to differentiate between benign and malignant lesions based on the six imaging parameters and patient age. Ten-fold cross-validation was applied to prevent overfitting. Overall diagnostic accuracy and decision rules (rule-out criteria) to accurately exclude malignancy were evaluated. Results were integrated into a web application and published online. Malignant lesions were present in 107 patients (60.8%). Imaging features showed excellent interobserver variability (ICC: 0.81-0.98) with variable diagnostic accuracy (AUC: 0.65-0.82). Overall performance of the ML algorithm was high (AUC = 90.1%; BI-RADS IV: AUC = 91.6%). The ML algorithm provided decision rules to accurately rule-out malignancy with a false negative rate <1% in 31.3% of the BI-RADS IV cases. Thus, integration of ML into MRI interpretation can provide objective and accurate decision rules for the management of suspicious breast masses, and could help to reduce the number of potentially unnecessary biopsies.

Implementation of machine learning into clinical breast MRI: Potential for objective and accurate decision-making in suspicious breast masses.

This study aimed to develop a tool for the classification of masses in breast MRI, based on ultrafast TWIST-VIBE Dixon (TVD) dynamic sequences combined with DWI. TVD sequences allow to abbreviate breast MRI protocols, but provide kinetic information only on the contrast wash-in, and because of the lack of the wash-out kinetics, their diagnostic value might be hampered. A special focus of this study was thus to maintain high diagnostic accuracy in lesion classification.

Combination of an ultrafast TWIST-VIBE Dixon sequence protocol and diffusion-weighted imaging into an accurate easily applicable classification tool for masses in breast MRI.

Integration of medical imaging into preclinical anatomy courses is already underway in many medical schools. However, interpretation of two-dimensional grayscale images is difficult and conventional volume rendering techniques provide only images of limited quality. In this regard, a more photorealistic visualization provided by Cinematic Rendering (CR) may be more suitable for anatomical education. A randomized, two-period crossover study was conducted from July to December 2018, at the University Hospital of Erlangen, Germany to compare CR and conventional computed tomography (CT) imaging for speed and comprehension of anatomy. Sixteen students were randomized into two assessment sequences. During each assessment period, participants had to answer 15 anatomy-related questions that were divided into three categories: parenchymal, musculoskeletal, and vascular anatomy. After a washout period of 14 days, assessments were crossed over to the respective second reconstruction technique. The mean interperiod differences for the time to answer differed significantly between the CR-CT sequence (-204.21 ± 156.0 seconds) and the CT-CR sequence (243.33 ± 113.83 seconds; P < 0.001). Overall time reduction by CR was 65.56%. Cinematic Rendering visualization of musculoskeletal and vascular anatomy was higher rated compared to CT visualization (P < 0.001 and P = 0.003), whereas CT visualization of parenchymal anatomy received a higher scoring than CR visualization (P < 0.001). No carryover effects were observed. A questionnaire revealed that students consider CR to be beneficial for medical education. These results suggest that CR has a potential to enhance knowledge acquisition and transfer from medical imaging data in medical education.

Article	Year
Advanced Modeled Iterative Reconstruction (ADMIRE) Facilitates Radiation Dose Reduction in Abdominal CT. Academic radiology\| PubMed ID: 29500115	2018
Prediction of early metastatic disease in experimental breast cancer bone metastasis by combining PET/CT and MRI parameters to a Model-Averaged Neural Network. Bone\| PubMed ID: 30445200	2019
Implementation of machine learning into clinical breast MRI: Potential for objective and accurate decision-making in suspicious breast masses. PloS one\| PubMed ID: 31999755	2020
Combination of an ultrafast TWIST-VIBE Dixon sequence protocol and diffusion-weighted imaging into an accurate easily applicable classification tool for masses in breast MRI. European radiology\| PubMed ID: 32002644	2020
Cinematic Rendering in Anatomy: A Crossover Study Comparing a Novel 3D Reconstruction Technique to Conventional Computed Tomography. Anatomical sciences education\| PubMed ID: 32521121	2020

Stephan Ellmann

Biography

JoVE Journal Publications

Other Publications