CuraCloud scientists presented Simultaneous Classification and Segmentation of Intracranial Hemorrhage Using a Fully Convolutional Neural Network at the 2020 17th IEEE International Symposium on Biomedical Imaging (ISBI). ISBI is a scientific conference focused on mathematical, algorithmic, and computational aspects of biological and biomedical imaging. Each year, the conference brings together leading imaging scientists and professionals from around the world.
In this blog post, we will provide a brief summary on how our scientists developed and tested a multi-task fully convolutional neural network, ICHNet, for the simultaneous detection, classification, and segmentation of ICH.
Intracranial hemorrhage (ICH) is a critical condition in which bleeding occurs within the cellular tissue and the spaces within the membranes surrounding the brain. When diagnosing ICH, accurate detection, subtype classification, and volume quantification are essential to saving patients’ lives and their subsequent recovery. Previous studies have applied deep learning for ICH analysis, but typically tackle the tasks in a separate manner without utilizing information sharing. Our scientists have proposed a multi-task fully convolutional network, ICHNet, for the simultaneous detection, classification, and segmentation of ICH. It contains a shared encoder to extract features for both classification (cls) and segmentation (seg) tasks, and utilizes a convolutional long short-term memory (ConvLSTM) module to capture the sequential information embedded in consecutive slices. We evaluated the performance of our proposed architecture to demonstrate that it improves the performance of both classification and segmentation tasks compared with single-task and baseline models.
To evaluate the performance of ICHNet, we used a total of 1,176 head CT scans from three participating hospitals, with 581 ICH patients and 595 normal subjects. Our scientists divided the dataset into a random sample of 706 subjects for training, 235 for validation, and 235 for testing. We used five metrics including, accuracy, sensitivity, specificity, F1 score, and area under the curve (AUC) to evaluate models’ performance for the cls tasks, and used dice coefficient as the evaluation metric for the seg task.
Figure 1. Visualizations of ICHNet segmentation results. Red denotes hemorrhages, and green denotes normal brain tissue. 3D ground truth masks (left) and segmentation results (right).
ICHNet generally outperformed the baseline models (3D/2D ResNet18 for cls and 3D U-Net for seg) as well as single-task ICHNet models (ICHNetcls and ICHNetseg) across all metrics. In addition, the single-task ICHNet model’s performance was notably better than the corresponding baseline model, indicating that ConvLSTM module can be a more successful approach in capturing sequential information than directly utilizing 3D convolution. In addition to performance enhancement, our proposed architecture offers more flexibility in practical use due to its compatibility with different levels of annotations.