Valeria Romeo1, Carlo Ricciardi1, Renato Cuocolo2, Arnaldo Stanzione1, Francesco Verde1, Laura Sarno3, Giovanni Improta4, Pier Paolo Mainenti5, Maria D'Armiento1, Arturo Brunetti1, Simone Maurea1. 1. University of Naples "Federico II", Department of Advanced Biomedical Sciences, Naples, Italy. 2. University of Naples "Federico II", Department of Advanced Biomedical Sciences, Naples, Italy. Electronic address: renato.cuocolo@unina.it. 3. University of Naples "Federico II", Department of Neuroscience, Reproductive and Dentistry Sciences, Naples, Italy. 4. University of Naples "Federico II", Department of Public Health, Naples, Italy. 5. Institute of Biostructures and Bioimaging of the National Council of Research (CNR), Naples, Italy.
Abstract
PURPOSE: To evaluate whether a machine learning (ML) analysis employing MRI-derived texture analysis (TA) features could be useful in assessing the presence of placenta accreta spectrum (PAS) in patients with placenta previa (PP). The hypothesis is that TA features may reflect histological abnormalities underlying PAS in patients with PP thus helping in differentiating positive from negative cases. MATERIALS AND METHODS: Pre-operative MRI examinations of 64 patients with PP of which 20 positive (12 accreta, 7 increta and 1 percreta) and 44 negative for PAS were retrospectively selected. Multiple (n = 3) rounded regions of interest (ROIs) were manually positioned on sagittal or coronal T2-weighted images over homogeneous placental tissue close to the placental-myometrial interface for each patient to extract TA features. After balancing the dataset with the Synthetic Minority Over-sampling Technique, training and testing sets were obtained using Hold-out with a 75/25% split. Different algorithms were applied on the training set using the wrapper method, which looks for the best combination of features based on the optimization of a heuristic function in order to get the highest accuracy, and a 10-fold Cross-validation. The accuracy of the best models was also assessed on the test set. Histology was used as the standard of reference. RESULTS: A total of 192 ROIs were positioned and a ROI-based analysis was then conducted. Among the different algorithms, k-nearest neighbors obtained the highest accuracy (98.1%), precision (98.7%), sensitivity (97.5%) and specificity (98.7%) while exploiting the lowest number of features (n = 26); conversely, the Naïve Bayes algorithm got the lowest scores showing an accuracy of 80.5%. CONCLUSION: ML analysis using MRI-derived TA features could be a feasible tool in the identification of placental tissue abnormalities underlying PAS in patients with PP. This approach might represent an additional tool in the clinical practice, thus expanding the application field of artificial intelligence to medical images.
PURPOSE: To evaluate whether a machine learning (ML) analysis employing MRI-derived texture analysis (TA) features could be useful in assessing the presence of placenta accreta spectrum (PAS) in patients with placenta previa (PP). The hypothesis is that TA features may reflect histological abnormalities underlying PAS in patients with PP thus helping in differentiating positive from negative cases. MATERIALS AND METHODS: Pre-operative MRI examinations of 64 patients with PP of which 20 positive (12 accreta, 7 increta and 1 percreta) and 44 negative for PAS were retrospectively selected. Multiple (n = 3) rounded regions of interest (ROIs) were manually positioned on sagittal or coronal T2-weighted images over homogeneous placental tissue close to the placental-myometrial interface for each patient to extract TA features. After balancing the dataset with the Synthetic Minority Over-sampling Technique, training and testing sets were obtained using Hold-out with a 75/25% split. Different algorithms were applied on the training set using the wrapper method, which looks for the best combination of features based on the optimization of a heuristic function in order to get the highest accuracy, and a 10-fold Cross-validation. The accuracy of the best models was also assessed on the test set. Histology was used as the standard of reference. RESULTS: A total of 192 ROIs were positioned and a ROI-based analysis was then conducted. Among the different algorithms, k-nearest neighbors obtained the highest accuracy (98.1%), precision (98.7%), sensitivity (97.5%) and specificity (98.7%) while exploiting the lowest number of features (n = 26); conversely, the Naïve Bayes algorithm got the lowest scores showing an accuracy of 80.5%. CONCLUSION:ML analysis using MRI-derived TA features could be a feasible tool in the identification of placental tissue abnormalities underlying PAS in patients with PP. This approach might represent an additional tool in the clinical practice, thus expanding the application field of artificial intelligence to medical images.
Authors: Valeria Romeo; Francesco Verde; Laura Sarno; Sonia Migliorini; Mario Petretta; Pier Paolo Mainenti; Maria D'Armiento; Maurizio Guida; Arturo Brunetti; Simone Maurea Journal: Radiol Med Date: 2021-06-22 Impact factor: 3.469
Authors: Christopher D Nguyen; Ana Correia-Branco; Nimish Adhikari; Ezgi Mercan; Srivalleesha Mallidi; Mary C Wallingford Journal: EMJ Radiol Date: 2020-09
Authors: Lorenzo Ugga; Renato Cuocolo; Domenico Solari; Elia Guadagno; Alessandra D'Amico; Teresa Somma; Paolo Cappabianca; Maria Laura Del Basso de Caro; Luigi Maria Cavallo; Arturo Brunetti Journal: Neuroradiology Date: 2019-08-02 Impact factor: 2.804
Authors: Allan F F Alves; Sérgio A Souza; Raul L Ruiz; Tarcísio A Reis; Agláia M G Ximenes; Erica N Hasimoto; Rodrigo P S Lima; José Ricardo A Miranda; Diana R Pina Journal: Phys Eng Sci Med Date: 2021-03-17