Model uczenia maszynowego oparty na rezonansie magnetycznym do rozróżnienia stwardnienia hipokampu typu 1 (ILAE) od samego zwłóknienia bez stwardnienia w padaczce płata skroniowego
MRI-based machine learning model to distinguish hippocampal sclerosis (HS) ILAE type 1 and no HS gliosis only in medial temporal lobe epilepsy
W skrócie
Naukowcy opracowali narzędzie do analizy zdjęć mózgu (rezonans magnetyczny), które potrafi rozróżnić dwa różne rodzaje zmian w hipokampie u pacjentów z oporną na leki padaczką płata skroniowego. Model ten osiągnął bardzo dobrą dokładność (82% specyficzność i 80% czułość) w odróżnieniu zwłóknienia od stwardnienia hipokampu. Ta umiejętność rozróżnienia tych zmian może pomóc lekarzom w wyborze odpowiedniego leczenia i prognozowaniu wyników operacji u pacjentów z padaczką.
Oryginalny abstract (angielski)
PURPOSE: Despite recent advances in preoperative work-up of drug resistant medial temporal lobe epilepsy (MTLE), predicting post-surgical seizure and memory outcomes remains challenging. Differentiating between "no hippocampal sclerosis Gliosis Only" [i.e. gliosis only; no pyramidal neuronal loss] and "HS ILAE Type 1" is essential as these two are distinct histopathological entities in MTLE with different clinical characteristics. In this study, we developed a Random Forest model with 10-fold cross validation using previously introduced qualitative and clinical MRI-based HS grading system to differentiate between "no HS Gliosis Only" and "HS ILAE Type 1". MATERIALS AND METHODS: 61 patients with drug resistant MTLE (mean age [SD], 39.5 [12.6] years; 36 [41.0%] male) underwent our routine institutional epilepsy protocol MRI on either a 1.5 T (36 patients) or 3 T (25 patients) scanner prior to temporal lobectomy or selective amygdalohippocampectomy. Histopathologic diagnosis of the resected hippocampi was assigned by a neuropathologist and divided into 3 categories: no HS Gliosis Only, HS ILAE 1 Total Sclerosis, and HS ILAE 1 Classic Sclerosis. Four experienced academic neuroradiologists arrived at consensus in grading the hippocampi of the patients. RESULTS: The HC total score demonstrated an excellent interrater agreement with ICC(3,1) of 1 95% CI (1, 1) and 0.98 [95% CI (0.94, 0.99)] for left and right hippocampus, respectively. Random Forest with 10-fold cross validation can achieve sensitivity of 80% [95% CI (55%-100%)], specificity of 82% [95% CI (72%-93%)], Positive Predictive Value (PPV) of 47% [95% CI (23%-71%)], and Negative Predictive Value (NPV) of 95% [95% CI (89%-100%)] respectively, for detecting "No HS Gliosis Only". A single Classification Tree revealed 2 as the diagnosis cut point for "HC total score". CONCLUSION: The distinction between "No HS Gliosis Only" and "HS ILAE Type 1" can be achieved by visual evaluation of clinical epilepsy protocol brain MRI. The ability to distinguish "No HS Gliosis Only" versus "HS ILAE Type 1" through visual evaluation of clinical epilepsy brain MRI and simple, interpretable decision tree can guide clinicians in prognosis, therapeutic decisions, and monitoring of disease progression, improving patient outcomes in MTLE management.