Classification of Myelin Oligodendrocyte Glycoprotein Antibody-Related Disease and Its Mimicking Acute Demyelinating Syndromes in Children Using MRI-Based Radiomics: From Lesion to Subject

Acad Radiol. 2024 May;31(5):2085-2096. doi: 10.1016/j.acra.2023.11.011. Epub 2023 Nov 25.

Abstract

Rationale and objectives: To develop MRI-based radiomics models from the lesion level to the subject level and assess their value for differentiating myelin oligodendrocyte glycoprotein antibody-related disease (MOGAD) from non-MOGAD acute demyelinating syndromes in pediatrics.

Materials and methods: 66 MOGAD and 66 non-MOGAD children were assigned to the training set (36/35), internal test set (14/16), and external test set (16/15), respectively. At the lesion level, five single-sequence models were developed alongside a fusion model (combining these five sequences). The radiomics features of each lesion were quantified as the lesion-level radscore (LRS) using the best-performing model. Subsequently, a lesion-typing function was employed to classify lesions into two types (MOGAD-like or non-MOGAD-like), and the average LRS of the predominant type lesions in each subject was considered as the subject-level radscore (SRS). Based on SRS, a subject-level model was established and compared to both clinical models and radiologists' assessments.

Results: At the lesion level, the fusion model outperformed the five single-sequence models in distinguishing MOGAD and non-MOGAD lesions (0.867 and 0.810 of area under the curve [AUC] in internal and external testing, respectively). At the subject level, the SRS model showed superior performance (0.844 and 0.846 of AUC in internal and external testing, respectively) compared to clinical models and radiologists' assessments for distinguishing MOGAD and non-MOGAD.

Conclusion: MRI-based radiomics models have potential clinical value for identifying MOGAD from non-MOGAD. The fusion model and SRS model can distinguish between MOGAD and non-MOGAD at the lesion level and subject level, respectively, providing a differential diagnosis method for these two diseases.

Keywords: Machine Learning; Multi-Sequence MRI; Myelin Oligodendrocyte Glycoprotein Antibody-related Disease; Pediatric; Radiomics.

MeSH terms

  • Adolescent
  • Autoantibodies
  • Child
  • Child, Preschool
  • Demyelinating Autoimmune Diseases, CNS / diagnostic imaging
  • Demyelinating Autoimmune Diseases, CNS / immunology
  • Demyelinating Diseases* / diagnostic imaging
  • Diagnosis, Differential
  • Female
  • Humans
  • Infant
  • Magnetic Resonance Imaging* / methods
  • Male
  • Myelin-Oligodendrocyte Glycoprotein* / immunology
  • Radiomics
  • Retrospective Studies

Substances

  • Myelin-Oligodendrocyte Glycoprotein
  • Autoantibodies