A Lightweight Deep-Feature and Agentic AI Fusion Framework for Automated Intracranial Hemorrhage Diagnosis from CT Images

Abstract

The correct identification of hematomas on the CT scan of the head is essential for the timely treatment of patients. However, the process of interpreting such images manually entails inter-reader variation. Computer-aided detection (CAD) provides reliable ways to increase the accuracy of medical imaging by introducing a higher level of standardization and automation in the analysis process. In this paper, we propose a hybrid solution for the classification of benign and malignant CT images of the head using a combination of deep feature extractiqon and machine learning algorithms. Our pipeline includes preprocessing and data augmentation steps for the normalization of input images. Deep convolutional neural networks, including esNet34, SqueezeNet, and MobileNetV2, were trained as feature encoders to extract discriminatory representations that can be used to train classification models like LightGBM, Random Forest, and XGBoost. Evaluation of the classification models' performance involved calculating their accuracy, precision, recall, specificity, F1-score, and ROC curve area under the curve. For all nine experiments performed, the MobileNetV2–LightGBM pipeline demonstrated the best performance, achieving 97% accuracy, precision, recall, specificity, and ROC curve AUC of 99.59%. Moreover, since our solution relies on modular training procedures, it can be easily combined with agentic artificial intelligence elements for autonomous operation.