Lesion size of early cerebral infarction on a Non-Contrast CT influences detection ability in Cascade Mask Region-Convolutional neural networks

[Display omitted] •Early brain infarction size on a NCCT can change Mask R-CNN’s performance.•Mask R-CNN can well detect relatively big early infarctions on NCCT.•To well detect smaller early infarctions on NCCT is the future goal. Currently, convolutional neural network (CNN) methods for early non-contrast brain CT (NCCT) infarction detection have limited accuracy. We suspect this may be due to the significant scale variance of lesion size, which would result in small infarctions not being properly detected. Patients’ early NCCT infarction signs were labeled and divided into training/validation and test datasets. The 1st Cascade Mask Region-Convolutional Neural Network (Cascade Mask R-CNN) model (1st model) was train/validated. The test dataset, with added normal controls, was used to evaluate the performances. We further divided the lesions into large and small ones by statistics. We filtered the test dataset’s ‘small’ lesions and re-examined the 1st model. Finally, we filtered the ‘small’ lesions of the training/validation datasets to train/validate the 2nd Cascade Mask R-CNN (2nd model), and the performances were examined by the filtered test dataset. In total, 266 ischemic stroke patients (Male/Female 163/103, age 69.9 ± 14.1 years) were included, and the duration from last known well to NCCT was 28.5 ± 39.3 h. Two hundred twenty patients were placed in training/validation datasets and 46 were assigned to the test dataset. Twenty normal control cases (Male/Female 14/6, age 68.7 ± 10.8 years) were added into the test dataset. The results were Sensitivity: 74.0%; Specificity: 91.8%; Accuracy: 86.1%; Precision: 80.9%; and F1-score: 77.3%. Based on the histogram, large lesions were > 2,012 pixels (4.8 cm2), and small lesions were ≤ 2,012 pixels. The 1st model’s performances by the filtered test dataset were Sensitivity: 96.9%; Specificity: 92.9%; Accuracy: 93.6%; Precision: 74.1%; and F1-score: 84.0%. The 2nd model’s performances were Sensitivity: 97.1%; Specificity: 96.4%; Accuracy: 96.5%; Precision: 85.7%; and F1-score: 91.0%. The target objects in most benchmark databases do not have the property of significant scale variance, which means commonly developed AI models focus on the issue of translational variance. The conventional CNN-based AI models do have the property of translation equivariance. The impact of scale variance has rarely been explored in real applications. For patients with ischemic stroke, the lesion size may vary significantly. Experi