Synergizing concatenated Efficientnetb3 and Inceptionv3 with SMOTE and Monte Carlo dropout for intracerebral hemorrhage detection and classification

Intracerebral hemorrhage (ICH) is an acute condition that threatens life urgently needs to be diagnosed quickly, and accurately, using medical imaging (primarily computed tomography (CT) scans). Nonetheless, the ICH dataset class imbalance, in which the great majority of cases are normal or non-hemorrhage causes significant difficulty for deep learning models to recognize the minority hemorrhage classes. In this study, we propose a novel hybrid deep learning model that is integrated with EfficientNet-B3 and InceptionV3 incorporating SMOTE and Monte Carlo Dropout for ICH detection and ICH subtypes classification. We tackle class imbalance by applying Synthetic Minority Over-sampling Technique (SMOTE) that produces synthetic lines from the minority classes, resulting in better generalization and better results in all classes. During feed forward we also used Monte Carlo Dropout to quantify metrics about model uncertainty, reducing model uncertainty and making predictions more robust and reliable for ambiguous cases. The model was trained and evaluated on a dataset consisting of 13,000 CT images, categorized into four classes: deep hemorrhage, no hemorrhage, lobar hemorrhage, and subtentorial hemorrhage. When compared without SMOTE, the model was able to hit an accuracy of 93%, but suffered poor performance for minority classes. Nevertheless, when SMOTE is applied to the model, accuracy results as high as 98% are recorded. F1-scores for minority classes (“yes/Lobar” and “yes/Subtentorial”) increased from 0.47 and 0.78 to 0.98 and 0.88, respectively, and the precision and recall for them also increased substantially. The SMOTE and Monte Carlo Dropout benefits were further confirmed by the uncertainty entropy analysis, which showed reductions in prediction uncertainty for minority classes.