Automatic stroke classification: Domain knowledge injection augmented transfer learning approach

Abstract

Aim: To build an artificial intelligence model to classify stroke into ischemic or hemorrhagic classes using the labeled stroke computer tomography (CT) slices that were shared in the 2021 Teknofest artificial intelligence in health competition. Methods: We developed a set of methods that can inject domain knowledge into the models to provide a more refined search space for the model for better performance. We used pre-trained MobileNet and EfficientNet architectures and fine-tuned them for our 2-class output model. We discarded impossible pixel values and pixel spatial locations to provide a space that was conditioned into only possible spatial locations and signal values using our knowledge of brain anatomy, stroke pathology, and imaging. Results: With the dataset which we just used [0-1] normalization and adjusted the input dimension into 224*224, accuracy values of 0.74 with adapted MobileNetV2 and 0.72 with adapted EfficentNetB0 were obtained in the group without further pre-processing. In the data transformation group where bone structures were removed and pixel values were restricted by eliminating impossible values, an accuracy level of 0.91 with MobileNetV2 and 0.88 with EfficientNetB0 at test time were achieved. Conclusion: In conclusion, CT-based slice prediction of mechanism of stroke as ischemic or hemorrhagic was achieved with high accuracy by integrating human knowledge into the pre-trained off-the-shelf models which was promising to shorten the time of the triage of stroke patients which can potentially improve stroke patient outcomes.

Ähnliche Arbeiten