Utilizzo delle nuove tecniche di Risonanza Magnetica muscolare nelle malattie neuromuscolari.

Background: In the context of neuromuscular diseases, utility of muscle Magnetic Resonance Imaging (MRI) in the diagnostic work-up and in monitoring the progression has been already assessed and proves useful in identifying different pattern of muscle involvement. The aims of this study are (1) deepen the current knowledge on muscle MRI applied to Distal Myopathies (DMs); (2) compare a large DM cohort with literature data; (3) validate quantitative mapping MR sequences, for clinical use; (4) pursue Artificial Intelligence (AI) approaches, to recognize involved muscles. Methods: (1) A systematic literature review focused on muscle MR in DMs was carried out. (2) Among the Italian Network of Distal Myopathies, all available muscle MR examinations were reviewed, blinded to clinical data, to analyze patterns of muscle involvement. Data from patients with pathogenetic mutations in myotilin (MYOT) gene were collected, to provide a full retrospective description of functional status and disease progression. (3) T1/T2 mapping values from muscle, with a standardized acquisition protocol (Siemens 1.5T) in a large homogeneous cohort for sex and age, were prospectively assessed; correlations with population physiological variables were also explored. (4) Unsupervised (Self-Organizing Map, SOM) and Supervised (Convolutional Neural Networks, CNN) Machine Learning methods were applied, to discriminate healthy and affected muscles at MRI. The SOM was "trained" to characterize the radiomic data extracted from a homogeneous population muscle MRIs; the output was compared with the "human" radiological evaluation. CNN was applied on labeled MR images; its performance and different explainable methods were reported. Results: (1) Most reports on muscle MRI in DMs rely on small series of patients, and often are lacking a systematic assessment; moreover, only for some disorders are there studies reporting the imaging characteristics. Consistently, pattern-based approaches to identify the underlying DM genotype have yielded low sensitivity and specificity. (2) 102 DM patients were evaluated. Typical patterns of patients with genetic diagnosis (80%) have been described for phenotype-genotype correlation; muscle imaging in patients without genetic diagnosis was also characterized. 29 subjects carrying MYOT gene mutations have been enrolled: our results show a slowly progressive muscle weakness often with onset in the distal posterior compartment of the legs and spreading to the anterior compartment and to proximal muscles in the lower limbs. To date, this is the largest MYOT population, compared to literature data. (3) Muscle mapping reference values stratified by sex and age were obtained from 50 healthy subjects using MOLLI and T2p-SSFP sequences, for T1/T2 mapping respectively. This optimized protocol can improve muscle tissue characterization on myopathies. In our knowledge, no systematic studies are available about the normal muscle mapping values. (4) The trained SOM was able to recognize muscles altered by the presence of oedema and fibrofatty replacement. Two different CNN setup were developed: both showed very high accuracy (>95%) and low misdiagnoses rate (2-4%). Explainable AI methods depict areas that contributed to the network’s diagnosis. Different techniques will be implemented to improve the network performance. Conclusion: Neuromuscular diseases still represent a challenge for diagnostic imaging. Structured phenotype-genotype evaluations on preferably large cohorts may improve the diagnostic accuracy. A standardization of MR protocols together with quantitative imaging, may allow a precise monitoring and comparison between patients, leading new diagnostic-therapeutic objectives. With this purpose, different AI approaches could be developed, with multidisciplinary experts’ contribution. Further studies are needed to discriminate different disorders and stratify disease progression through MRI tools.