Source-linked AI summary
Accurate and Efficient Intracranial Hemorrhage Detection and Subtype Classification in 3D CT Scans with Convolutional and Long Short-Term Memory Neural Networks
Mihail Burduja, Radu Tudor Ionescu, Nicolae Verga
TL;DR
Rapid intracranial hemorrhage diagnosis matters, but manual CT review is time-consuming. The paper combines slice-level CNN features with scan-level LSTM processing and feature selection; it reports a 0.04989 challenge loss and radiologist-assessed accuracy on par with specialized doctors.
Problem
Manual analysis of brain CT scans is complicated and time-consuming, while intracranial hemorrhage requires urgent diagnosis.
Source: §1 · p. 2Method
A CNN extracts features from individual CT slices, feature selection reduces them, and an LSTM processes the embeddings across a scan.
Source: §1 · p. 2Results
0.04989 weighted mean log loss placed the system 27th (top 2%) among 1,345 challenge participants, and radiologists assessed its accuracy as on par with specialized doctors.
Source: §1 · p. 2Takeaways & Limitations
The authors consider the system a viable option when fast diagnosis or a second opinion on intracranial hemorrhage detection is needed.
Source: §1 · p. 3The system can predict a subtype correctly without considering all locations for that subtype, which is not optimal for visualizing all affected locations.
Source: §6 · p. 18
Abstract
from arXiv · showhide
In this paper, we present our system for the RSNA Intracranial Hemorrhage Detection challenge. The proposed system is based on a lightweight deep neural network architecture composed of a convolutional neural network (CNN) that takes as input individual CT slices, and a Long Short-Term Memory (LSTM) network that takes as input feature embeddings provided by the CNN. For efficient processing, we consider various feature selection methods to produce a subset of useful CNN features for the LSTM. Furthermore, we reduce the CT slices by a factor of 2x, allowing ourselves to train the model faster. Even if our model is designed to balance speed and accuracy, we report a weighted mean log loss of 0.04989 on the final test set, which places us in the top 30 ranking (2%) from a total of 1345 participants. Although our computing infrastructure does not allow it, processing CT slices at their original scale is likely to improve performance. In order to enable others to reproduce our results, we provide our code as open source at https://github.com/warchildmd/ihd. After the challenge, we conducted a subjective intracranial hemorrhage detection assessment by radiologists, indicating that the performance of our deep model is on par with that of doctors specialized in reading CT scans. Another contribution of our work is to integrate Grad-CAM visualizations in our system, providing useful explanations for its predictions. We therefore consider our system as a viable option when a fast diagnosis or a second opinion on intracranial hemorrhage detection are needed.
1. Introduction
Intracranial hemorrhage requires urgent diagnosis, but manual CT review is time-consuming. The paper presents an efficient CNN–LSTM system, reports strong challenge and radiologist-assessment results, and offers visual explanations and open-source code.
1. Introduction
- Manual CT analysis by radiologists is complicated and time-consuming, delaying intervention for a life-threatening condition that requires urgent diagnosis.The introduction cites roughly 10% of US strokes as intracranial hemorrhages and estimates 30,000 related deaths annually.Source: §1 · pp. 1–2
- The CNN processes individual CT slices, while an LSTM processes their feature embeddings across a scan; feature selection and 2× downscaling target computational efficiency.The system also concatenates CNN predictions with LSTM feature vectors, which the authors say boosts accuracy with negligible computational overhead.Source: §1 · p. 2
- 0.04989 weighted mean log loss placed the system 27th (top 2%) among 1,345 RSNA challenge participants; radiologists judged its accuracy on par with specialized doctors.The authors note that the challenge score alone does not indicate performance relative to highly trained radiologists.Source: §1 · p. 2
- The paper presents Grad-CAM explanations, radiologist interpretations of difficult cases, and publicly available code, and concludes the system may aid fast diagnosis or a second opinion.The authors list joint CNN–recurrent modeling, human evaluation, reproducible code, and model-decision interpretations as contributions.Source: §1 · pp. 2–3
2. Related Work
Prior work spans conventional machine learning, CNNs, segmentation networks, and CNN–recurrent models for intracranial hemorrhage. This paper’s related-work discussion distinguishes its scan-level subtype-classification approach from these lines of work.
2. Related Work
- The review covers deep-learning and conventional approaches; it characterizes several statistical or standard machine-learning methods as requiring manual parameter adjustment and monitoring for large datasets.Examples include fuzzy C-means, thresholding, superpixels, and level-set methods.Source: §2 · p. 3
- Earlier CNN detection systems processed slices or scans in varied ways, while this work feeds penultimate-layer CNN activations into an LSTM that processes whole scans.The authors say this scan-level processing can correct erratic slice-level predictions.Source: §2 · p. 4
- FCN approaches primarily target segmentation, whereas this work targets detection with a more efficient 3D-scan architecture that does not require 3D convolutions.The authors note their CNN–LSTM is not suitable for segmentation, though Grad-CAM visualizations can be interpreted as rough segmentation maps.Source: §2 · p. 4
- Unlike related CNN–LSTM approaches [5], the proposed method classifies five hemorrhage subtypes; compared with Ye et al., it performs detection and subtype classification in one pass using selected CNN features.The paper identifies Ye et al. as the only known similar design with five-subtype detection capability.Source: §2 · pp. 4–5
3. Materials & Methods
The method processes CT scans as ordered slice sequences: a pretrained CNN extracts slice features, feature selection reduces their size, and a bidirectional LSTM predicts at scan level. Preprocessing uses three tissue windows, augmentation, and a reduced image size; training uses multi-label loss and separate CNN and LSTM stages.
3. Materials & Methods
- The method trains a slice-level CNN, selects penultimate-layer features, then trains a scan-level bidirectional LSTM, with neighboring slices improving prediction over slice-level outputs.The CNN starts from pretrained ResNeXt-101 32×8d weights; the bidirectional LSTM is trained from scratch.Source: §3 · pp. 5–6
3.1. Data Preprocessing and Augmentation
- Each 512 × 512 CT slice is converted using brain, subdural, and soft-tissue HU windows into three channels, then resized to 256 × 256 × 3 and normalized.Augmentation uses horizontal flips, shifting, rotation, scaling, and brightness adjustment.Source: §3.1 · p. 6
3.2. Model and Feature Selection
- Feature selection reduces 2,048-dimensional CNN vectors for the LSTM, using Softmax-layer weights or PCA; the described experiments use PCA on 30,000 vectors and choose 120 or 192 components.The ResNeXt-101 32×8d model is pretrained and fine-tuned for slice-level hemorrhage prediction; its penultimate layer supplies the embeddings.Source: §3.2 · pp. 7, 9
- The three-layer bidirectional LSTM accepts variable-length sequences of slice embeddings and outputs 256-dimensional embeddings; CNN class probabilities are concatenated before final classification.The model jointly predicts hemorrhage presence and five subtype probabilities.Source: Figure 3 · p. 8, §3.2 · p. 9
3.3. Loss and Optimization
- The models jointly predict six independent probabilities and are trained with summed binary cross-entropy; the loss was chosen to match the challenge’s weighted-log-loss ranking metric.CNN and LSTM are trained separately with Adam, for three and four epochs respectively.Source: §3.3 · p. 10
4. Experiments and Results
The experiments evaluate hemorrhage detection on the RSNA 2019 Brain CT Hemorrhage data set, comparing CNN–LSTM variants and reporting official-test and validation results. The final system combines feature selection with reduced-resolution inputs to balance performance and computational cost.
4.1. Data Set
- The data set contains 25,272 CT examinations and 870,301 slices labeled for five hemorrhage subtypes; scans average about 35 slices.The official training set was split into 728,513 training slices and 24,290 validation slices; test labels were unavailable to participants.Source: §4.1 · p. 10, Table 1 · p. 11
4.2.1. Implementation Details
- Implementation used a single NVIDIA P100 GPU, 256 × 256 inputs, batches of 32 for CNN training, and top-120 or top-192 features for the bidirectional LSTM.CNN embeddings began at 2,048 features; training used mixed precision and Adam.Source: §4.2.1 · p. 11
4.2.2. Evaluation Metrics
- Evaluation used multi-label weighted mean log loss for slice-level challenge ranking and accuracy, AUC, sensitivity, and specificity for validation assessment.Because test labels were unavailable, the additional metrics were reported only on the validation set.Source: §4.2.2 · p. 12
4.3. Results on the Official Test Set
- ResNeXt-101 32×8d outperformed EfficientNet-B4 in the reported fine-tuning comparison, with a log loss of 0.06259, and was selected for subsequent experiments.EfficientNet-B4 was the largest EfficientNet architecture the authors could afford to train on their data set.Source: §4.3 · p. 12
- 0.04989 weighted mean log loss placed the final submission 27th of 1,345 challenge participants.The final system used PCA-selected 120-dimensional CNN features and concatenated CNN class probabilities with the LSTM features.Source: Table 2 · p. 12, §4.3 · p. 13
- Among considered methods, the authors chose PCA for final submissions; with the same bidirectional LSTM and 192 features, PCA achieved 0.05096 log loss, while 120-dimensional CNN features achieved 0.05022.Experiments without feature selection could not be performed on the authors’ computing infrastructure, so they did not establish performance without it.Source: §4.3 · p. 13
- Reducing inputs to 256 × 256 pixels and CNN embeddings to 120 features cut processing from 128 to 36 milliseconds per image.One training epoch took about 26 hours at full resolution versus 7 hours for the final model, which fit the platform’s 9-hour limit.Source: §4.3 · p. 13
4.4. Results on the Validation Set
- For EPH, IPH, and SAH, the bidirectional LSTM improved slice-level sensitivity by more than 10% over the plain ResNeXt-101 32×8d.EPH had the lowest sensitivity: 44.30% at slice level and 40.00% at scan level; the authors suggest its underrepresentation may explain this result.Source: §4.4 · p. 13
5. Assessment by Radiologists and Discussion
A held-out radiologist assessment found the models close to specialist performance, while Grad-CAM review of difficult or controversial cases revealed label disagreements and localization shortcomings.
5. Assessment by Radiologists and Discussion
- Models scored close to specialists #2 and #3, exceeded them for EPH and IPH, and the plain CNN reached 98% accuracy for generic ICH.The reported top accuracies were 100% for EPH and 92% for IPH.Source: §5 · p. 14, Table 5 · p. 15
- The comparison used 100 held-out CT scans, and doctors annotated them without access to ground-truth labels.The evaluation metrics were computed at scan level; some scans had multiple hemorrhage subtypes.Source: §5 · p. 14, Table 4 · p. 14, Table 5 · p. 15
- The doctors considered the visualizations useful and judged the system sufficiently accurate as a second opinion in daily practice.They identified emergency-room patients suspected of ICH as a potential use case.Source: §5 · pp. 15, 18
- Doctor #2 and the CNN shared 22 wrong labels; reassessment identified at least 25 incorrect ground-truth labels and another 5 disputable labels.The authors note that the shared errors suggested some ground-truth labels might be wrong.Source: §5 · p. 15
- Among 414 visualizations from difficult or controversial slices, doctors rated 54% correct, 19% partially correct, and 27% incorrect.The authors considered the incorrect-visualization share not high enough to represent a serious concern given the selected cases.Source: §5 · p. 16
- For generic ICH, training required detecting only one hemorrhage region, although doctors expected visualizations to indicate all regions; they recommended jointly reviewing maps for all detected types.The authors report that visualizations were generally strongest for IPH, IVH, and SAH, and weakest for generic ICH.Source: §5 · pp. 16–17
6. Conclusions
The joint ResNeXt–BiLSTM system reached specialist-comparable performance and placed 27th in the RSNA challenge, with a reduced computational footprint. Its subtype predictions can be correct without identifying every hemorrhage location, a limitation the authors aim to address with pixel-level annotations.
6. Conclusions
- 0.04989 weighted mean log loss earned 27th place (top 2%) among 1345 RSNA challenge participants, with performance comparable to trained specialists.The authors report these results alongside reducing CT slices from 512 × 512 to 256 × 256 pixels and selecting CNN features before the bidirectional LSTM.Source: §6 · p. 18
- The joint ResNeXt–BiLSTM approach outperformed a single ResNeXt model analyzing CT slices, and feature selection was effective for detection and subtype classification.The model reduces the computational footprint by halving slice dimensions and selecting CNN embeddings before the bidirectional LSTM.Source: §6 · p. 18
- Radiologists’ subjective assessment found the system useful for daily practice, providing predictions and visualizations for selected difficult or controversial slices.The paper reports an analysis of Grad-CAM visualizations.Source: §6 · p. 18
- Subtype predictions may be correct without identifying all locations of that hemorrhage subtype, limiting visualization of every affected region.The authors plan to incorporate pixel-level annotations, such as segments or bounding boxes, to penalize missed regions during training.Source: §6 · pp. 18–19
Abbreviations
The section lists abbreviations for organizations, imaging, neural-network methods, analysis tools, and hemorrhage subtypes.
Abbreviations
- RSNA, DICOM, CT, and HU denote the organization and imaging-related terms used in the manuscript.The listed expansions are Radiological Society of North America, Digital Imaging and Communications in Medicine, computed tomography, and Hounsfield unit.Source: Abbreviations · p. 19
- CNN, LSTM, RNN, and FCN identify the neural-network architectures and model components named in the paper.The list expands these as convolutional neural network, Long Short-Term Memory, recurrent neural network, and fully convolutional network.Source: Abbreviations · p. 19
- PCA, WC, WW, GPU, AMP, ROC, and AUC cover analysis, imaging settings, computing, and evaluation terms.The list also defines EPH, IPH, IVH, SAH, and SDH as hemorrhage subtype abbreviations.Source: Abbreviations · p. 19