Source-linked AI summary
Intracranial Hemorrhage Segmentation Using Deep Convolutional Model
Murtadha D. Hssayeni, M. S., Muayad S. Croock, Ph. D., Aymen Al-Ani, Ph. D., Hassan Falah Al-khafaji, M. D., Zakaria A. Yahya, M. D., Behnaz Ghoraani, Ph. D
TL;DR
Traumatic brain injury can cause life-threatening intracranial hemorrhage, whose CT diagnosis and localization rely heavily on experienced radiologists. This paper releases an annotated dataset and develops a U-Net fully convolutional network for automated ICH segmentation, achieving a Dice coefficient of 0.31 with 5-fold cross-validation.
Problem
ICH diagnosis and region localization on CT scans rely heavily on experienced radiologists, while public datasets for ICH segmentation are lacking.
Method
The study collected 82 traumatic-brain-injury CT scans with radiologist annotations and developed a U-Net fully convolutional network for slice-level ICH segmentation.
Results
A Dice coefficient of 0.31 was achieved for ICH segmentation using subject-based 5-fold cross-validation on 160 × 160 CT crops.
Takeaways & Limitations
The dataset is publicly available at PhysioNet, providing a benchmark for future ICH segmentation analysis and comparison.
Takeaways & Limitations
False positives near bones reduced Dice performance, and the model failed to localize small IPH regions in two subjects, so it was not suitable as a standalone method.
Abstract
from arXiv · showhide
Traumatic brain injuries could cause intracranial hemorrhage (ICH). ICH could lead to disability or death if it is not accurately diagnosed and treated in a time-sensitive procedure. The current clinical protocol to diagnose ICH is examining Computerized Tomography (CT) scans by radiologists to detect ICH and localize its regions. However, this process relies heavily on the availability of an experienced radiologist. In this paper, we designed a study protocol to collect a dataset of 82 CT scans of subjects with traumatic brain injury. Later, the ICH regions were manually delineated in each slice by a consensus decision of two radiologists. Recently, fully convolutional networks (FCN) have shown to be successful in medical image segmentation. We developed a deep FCN, called U-Net, to segment the ICH regions from the CT scans in a fully automated manner. The method achieved a Dice coefficient of 0.31 for the ICH segmentation based on 5-fold cross-validation. The dataset is publicly available online at PhysioNet repository for future analysis and comparison.
1. Introduction
Intracranial hemorrhage is a dangerous complication of traumatic brain injury that requires timely detection and localization. The paper motivates automated CT-based segmentation to reduce reliance on expert radiologists and addresses the lack of public segmentation datasets.
- ICH after traumatic brain injury can cause paralysis or death if untreated in a time-sensitive procedure.
- CT is preferred for initial ICH assessment because it is widely available and rapidly acquired.
- Radiologists inspect CT images to determine whether ICH occurred, identify its subtype, and localize its region.
- A fully convolutional network was developed to segment ICH regions in individual CT slices automatically.
- Publicly available datasets supported ICH subtype detection, but no public dataset existed for ICH segmentation.
2. Related work
Prior ICH research spans detection, subtype classification, and segmentation using traditional and deep learning methods. Although reported performance can be high, small datasets and difficult subtypes motivate publicly benchmarked segmentation methods that delineate all ICH subtypes.
- Most prior studies addressed binary ICH detection or multiclass subtype classification, while fewer targeted region segmentation.
- 98% sensitivity and 59% specificity were reported for one threshold-based ICH detector evaluated on 210 CT scans.
- 100% testing sensitivity, 92% specificity, and 91% accuracy were reported for a Bayesian decision method tested on 69 CT scans.
- Some deep learning approaches modeled spatial dependency across adjacent CT slices using random forests, recurrent networks, or slice interpolation.
- 99% slice-level sensitivity and specificity with an AUC of 1 were reported for one deep detection model, but subtype sensitivity was lower for SAH and EDH.
- Traditional approaches commonly required skull and noise removal, brain registration, and engineered features before segmentation.
- Deep segmentation methods used CNN, FCN, autoencoder-plus-active-contour, DRN, modified VGG16, and U-Net designs.
- The review identifies a remaining need for precise delineation of all ICH subtypes and motivates a public CT segmentation dataset with a U-Net implementation.
3. Dataset
The study assembled a retrospective traumatic-brain-injury CT dataset with radiologist-consensus ICH annotations, multiple image windows, and binary segmentation masks. It includes examples of several ICH sub-types and is publicly released for research.
- Data collection: 82 subjects underwent non-contrast CT annotation by two radiologists who reached consensus on ICH diagnosis, shape, and location.The dataset contains approximately 30 slices per CT scan.
- Data collection: CT slices were saved as 650x650 grayscale JPG images from brain and bone windows, alongside binary 650x650 JPG ICH masks.The collection workflow used a custom Matlab tool for slice navigation, annotation recording, and region delineation.
- Dataset composition: 36 of 82 subjects had ICH, comprising IVH, IPH, SAH, EDH, and SDH; one chronic ICH case was excluded.Some slices contained two or more ICH sub-types.
- Dataset composition: The ICH sub-type distribution is imbalanced, with most slices lacking ICH, IVH present in five subjects, and SDH present in four subjects.The imbalance should be considered when interpreting analyses using this dataset.
- Dataset release: The dataset was released in JPG and NIfTI formats through PhysioNet under a Creative Commons Attribution 4.0 International license.PhysioNet is described as a repository of freely available medical research data.
4. ICH Segmentation Using U-Net
The study applies U-Net, a fully convolutional network, to segment ICH regions in CT slices. Its architecture combines contracting and expansive paths, with each slice divided into 16 windows before segmentation.
- 4. ICH Segmentation Using U-Net: U-Net was investigated as a fully convolutional network for ICH segmentation, motivated by its effectiveness on small biomedical image datasets.The paper presents this as the first application of U-Net to ICH segmentation.
- 4. ICH Segmentation Using U-Net: Each CT slice is divided into 16 windows before being fed to U-Net for ICH segmentation.
- 4. ICH Segmentation Using U-Net: The architecture is symmetrical, pairing a contracting path with an expansive path for feature extraction and reconstruction.The contracting path uses convolutional blocks and max-pooling, while the expansive path upsamples feature maps.
- 4. ICH Segmentation Using U-Net: Skip connections concatenate cropped contracting-path features with upsampled expansive-path features to preserve fine-grained spatial information for precise localization.
- 4. ICH Segmentation Using U-Net: The network outputs an ICH probability for each pixel using a final 1 × 1 convolution with sigmoid activation and omits dense layers.The architecture contains 24 convolutional layers, four max-pooling layers, four upsampling layers, and four concatenations.
5. Experiments
The experiments compare threshold-based segmentation with U-Net using full CT slices and cropped windows. Jaccard index and Dice similarity coefficient quantify agreement with radiologist segmentation masks.
- 5. Experiments: Three experiments evaluated U-Net and compared it with a threshold-based method for ICH segmentation.The experiments varied the thresholding strategy and whether U-Net processed full slices or smaller crops.
- 5. Experiments: The threshold-based experiment used grid search to select lower and upper intensity thresholds for ICH regions.
- 5. Experiments: U-Net was evaluated using either full 640 × 640 CT slices or 160 × 160 crops.The crop-based setup was intended to address the small proportion of positive ICH pixels.
- 5. Experiments: Slice-level Jaccard index and Dice similarity coefficient measured how closely predicted segmentations matched ground-truth segmentations.The ground truth and U-Net segmentation regions are represented as R_ICH and R̂_ICH in the Dice formulation.
6. Results
The study used subject-based 5-fold cross-validation and addressed class imbalance through random undersampling of non-ICH slices or windows. The crop-based U-Net achieved a testing Dice coefficient of 0.31, but false positives and missed small hemorrhages remained.
- 6. Results: Subject-based 5-fold cross-validation trained, validated, and tested the model across all experiments.
- 6. Results: 36 of 82 subjects had ICH, yielding 318 ICH slices among 2,491 total slices, or fewer than 10% of images.The dataset included multiple ICH sub-types, with imbalanced representation across slices and subjects.
- 6. Results: Random undersampling reduced non-ICH slices or windows in the training data to address class imbalance.The crop-based experiment used 160 × 160 windows instead of only full 640 × 640 slices.
- 6. Results: 0.21 testing Jaccard index and 0.31 Dice coefficient were achieved by U-Net in the crop-based experiment.The same experiment reported slice-level sensitivity of 97.2% and specificity of 50.4%.
- 6. Results: The full-slice U-Net produced only black masks, whereas windowing improved the class-bias problem and yielded better cross-validated performance.
- 6. Results: The results included false-positive regions and a case where only part of an ending EDH region was segmented.
- 6. Results: Dice performance varied by ICH subtype, reaching 0.52 for SDH and averaging 0.35, 0.3, 0.28, and 0.23 for EDH, IVH, IPH, and SAH, respectively.The model failed to localize small IPH regions in two subjects.
7. Discussions
The study collected 82 CT scans and evaluated U-Net segmentation using full slices and 160 × 160 crops. Crop-based U-Net achieved a Dice coefficient of 0.31, with false positives and missed small hemorrhages limiting standalone use.
- 82 CT scans were collected, including 36 scans with detected ICH regions, for training and evaluating threshold-based and U-Net methods.
- U-Net was trained on full CT slices and on 160 × 160 overlapping crops with negative-class undersampling to address data imbalance.
- 0.31 Dice coefficient was achieved for ICH segmentation using U-Net trained on 160 × 160 crops, establishing a baseline for this dataset.
- Higher segmentation performance was reported by models trained and tested on larger datasets, including an average Dice coefficient of 0.85 in one study.
- False-positive segmentation was more prevalent near bones because their grayscale intensity resembled that of ICH regions.
- The model failed to localize small IPH regions in two subjects and was therefore characterized as assistive rather than standalone technology.
8. Conclusions
The paper contributes a publicly available dataset of 82 CT scans with ICH segmentation and a deep learning method evaluated by 5-fold cross-validation. It reports a Dice coefficient of 0.31 and frames automated screening as potentially useful where experts are unavailable.
- The paper collected 82 CT scans with ICH segmentation and made the dataset publicly available through PhysioNet.
- A deep learning method for ICH segmentation was assessed on the collected data using 5-fold cross-validation.
- 0.31 Dice coefficient was obtained for ICH segmentation and described as comparable to deep learning methods trained on small datasets.
- The paper includes a detailed review of methods for ICH detection, subtype classification, and segmentation.
Abbreviations
The manuscript defines abbreviations for imaging, injury, hemorrhage subtypes, neural-network methods, and evaluation terminology.
- CT denotes Computerized Tomography, TBI denotes Traumatic brain injury, and ICH denotes Intracranial hemorrhage.
- IVH, IPH, SAH, EDH, and SDH denote intraventricular, intraparenchymal, subarachnoid, epidural, and subdural hemorrhage, respectively.
- CNN, RNN, FCN, and LSTM denote convolutional, recurrent, fully convolutional, and long short-term memory neural networks.
- AUC denotes area under the ROC curve.