Artificial-Intelligence fosters the muscular MRI interpretation by the clinician: a proof-of-concept study

/by

Raffaele Conte 1, Gianluca Diodato 1, Maria Cristina Scudellari 1, Alessandro Tonacci 1, Stefano Roccella 2, Mario Milazzo 2, Guja Astrea 3, Filippo Maria Santorelli 3, Giulia Ricci 4, Gabriele Siciliano 4, Francesca Torri 4, Giacomo Aringhieri 4, Diego Lopergolo 5, Alessandro Malandrini 5, Sabrina Matà 6, Martina Sperti 6, Corrado Angelini 7, Francesco Sansone 1

1 Institute of Clinical Physiology, National Research Council of Italy (IFC-CNR), Pisa, Italy; 2 Biorobotics Institute, Sant’Anna School of Advanced Studies, Livorno, Italy; 3 IRCCS Stella Maris Foundation, Calambrone, Italy; 4 Pisa University Hospital, Pisa, Italy; 5 Siena University Hospital, Siena, Italy; 6 Careggi University Hospital, Firenze, Italy; 7 University of Padua, Padova, Italy

The present study aims at developing convolutional neural networks (CNNs) for automated segmentation of muscles in the thighs. Eighteen patients were retrospectively enrolled: 8 patients with neuromuscular alterations on MRI and 10 with unremarkable MRI. All patients performed MRI including dual echo, STIR sequences and DWI with b value set to 0 and 800.
A radiologist manually segmented thighs muscles with a different mask for each of the three muscle compartments (anterior, posterior and medial) on out-of-phase sequence. Manual segmentations were employed as groundtruth to train and validate two CNNs with different automatic segmentation UNet architectures, one for the thigh muscles together (UNet1) and one for the three muscular compartments, separately (UNet3). For each patient, the 2D slice at the middle third of the thighs was selected. Then, images of left and right thighs were divided into two separated datasets to double the sample size. Particularly, 15 thighs were used for training, 10 for validation, 11 for testing. The CNN performances were measured in terms of dice similarity coefficient (DSC).
UNet1 and UNet3 achieved an accuracy of 96% and 95%, respectively. UNet1 showed a DSC of 0.94, while UNet3 showed a mean DSC of 0.92, 0.80, and 0.90 for the segmentation of the anterior, medial, and lateral compartment, respectively.
Overall, UNet1 and UNet3 achieved an excellent performance in thighs muscles segmentation. This results paved the way for further, potentially effective AI-based models for quantification of neuromuscular alterations or radiomics features extraction, leading to a more practical, less time-consuming workflow.

Acknowledgment: The present work was supported by the project InGene 2.0, funded by the Tuscany Region under the Bando Salute 2018.