Comment on “Machine Learning for Early Detection of Hypoxic‑ischemic Brain Injury After Cardiac Arrest”

Abstract

With great interest, we have read the article by Mansour et al. [1], reporting on the use of deep transfer learning to identify early signs of hypoxic-ischemic brain injury (HIBI) on head computed tomography (HCT) scans. The authors report a very high accuracy (0.94) of their model with respect to the detection of HIBI signs on HCT scans performed within hours after the return of spontaneous circulation. The authors conclude that "Deep transfer learning reliably identifies HIBI in normal appearing findings on HCT performed within 3 h after ROSC in comatose survivors of a cardiac arrest" [1]. This interpretation is likely too optimistic.Deep learning networks show poor classification results and tend to be overfitted when trained on a very small data set [2]. A medical imaging data set of 54 HCT scans is a very small training data set. Further, we think that the following methodological issues could also contribute to overfitting in this study: (1) choice of the network, (2) the training pipeline (data augmentation, early stopping), and (3) principal component analysis (PCA) and repeated data usage.No justification was given for why a VGG19 network was chosen, although it has a significantly worse accuracy in the analysis of CT data than, for instance, ResNet-50 or . At the same time, it

Translational Applications of Neuroimaging

“…In a recent publication by Mansour et al (2022), machine and deep learning were utilized to identify patients who would exhibit radiologic evidence of apparent HIE on follow-up CT scans. Although this study demonstrates the potential of deep learning in detecting features that may not be visible to human raters, their proposed method included various significant limitations (i.e., high risk of overfitting due to small data set, questionable training pipeline and principal component analysis), which could result in partially erroneous results (Molinski et al, 2022). Our approach involved training our deep learning models from scratch, with direct class prediction as the output, without manual feature selection or additional machine learning modeling.…”

“…Another area with translational potential for neuroimaging is concerned with monitoring changes in brain structure following disease or interventions. Molinski et al (2024) used deep learning to detect hypoxic ischemic encephalopathy after cardiac arrest based on computer tomography (CT) images. Their overall sample comprised 168 CT images, of which about half (52.4%) showed radiological signs, according to expert ratings that were used as ground truth labels.…”

Deep learning-enabled detection of hypoxic–ischemic encephalopathy after cardiac arrest in CT scans: a comparative study of 2D and 3D approaches

“…In a recent publication by Mansour et al (2022) , machine and deep learning were utilized to identify patients who would exhibit radiologic evidence of apparent HIE on follow-up CT scans. Although this study demonstrates the potential of deep learning in detecting features that may not be visible to human raters, their proposed method included various significant limitations (i.e., high risk of overfitting due to small data set, questionable training pipeline and principal component analysis), which could result in partially erroneous results ( Molinski et al, 2022 ). Our approach involved training our deep learning models from scratch, with direct class prediction as the output, without manual feature selection or additional machine learning modeling.…”