Source-linked AI summary
RADNET: Radiologist Level Accuracy using Deep Learning for HEMORRHAGE detection in CT Scans
Monika Grewal, Muktabh Mayank Srivastava, Pulkit Kumar, Srikrishna Varadarajan
TL;DR
Brain hemorrhage detection from CT requires fast, accurate assessment that traditionally depends on time-consuming radiologist inspection. RADnet emulates radiologists by combining attention-based slice analysis with bidirectional LSTM modeling of neighboring slices, then achieves radiologist-comparable accuracy with higher recall than two radiologists. The method shows potential as an emergency diagnosis tool, but remains limited by a small test set and coverage of only hemorrhage rather than other severe brain conditions.
Problem
Automated hemorrhage detection must provide fast, accurate CT assessment comparable to radiologists because manual diagnosis is time-consuming and requires trained specialists.
Method
RADnet combines DenseNet, attention focused on hemorrhagic regions, bidirectional LSTM inter-slice modeling, and CT-level aggregation of slice predictions.
Results
81.82% CT-level accuracy was achieved, matching one radiologist; recall was higher than two radiologists and F1 score higher than two of three.
Takeaways & Limitations
RADnet demonstrates potential as an emergency diagnosis tool, with radiologist-comparable accuracy and increased sensitivity.
Takeaways & Limitations
The method was tested on a limited test set, and it does not address other equally severe brain conditions or replace radiologists.
Abstract
from arXiv · showhide
We describe a deep learning approach for automated brain hemorrhage detection from computed tomography (CT) scans. Our model emulates the procedure followed by radiologists to analyse a 3D CT scan in real-world. Similar to radiologists, the model sifts through 2D cross-sectional slices while paying close attention to potential hemorrhagic regions. Further, the model utilizes 3D context from neighboring slices to improve predictions at each slice and subsequently, aggregates the slice-level predictions to provide diagnosis at CT level. We refer to our proposed approach as Recurrent Attention DenseNet (RADnet) as it employs original DenseNet architecture along with adding the components of attention for slice level predictions and recurrent neural network layer for incorporating 3D context. The real-world performance of RADnet has been benchmarked against independent analysis performed by three senior radiologists for 77 brain CTs. RADnet demonstrates 81.82% hemorrhage prediction accuracy at CT level that is comparable to radiologists. Further, RADnet achieves higher recall than two of the three radiologists, which is remarkable.
1. INTRODUCTION
Brain hemorrhage diagnosis from CT is time-critical and traditionally depends on time-consuming radiologist inspection. RADnet combines slice-level analysis, 3D inter-slice context, and attention, then benchmarks performance against senior radiologists.
- CT-based hemorrhage diagnosis after traumatic brain injury is time-critical, while manual inspection requires trained radiologists and is time-consuming.
- RADnet models radiologists’ workflow by combining slice-level classification with sequence labeling to incorporate neighboring-slice context.
- The method combines DenseNet features, bidirectional LSTM spatial dependencies, and attention focused on hemorrhagic regions.
- The study benchmarks RADnet against three senior radiologists to assess whether automated inference reaches specialist-level reliability and sensitivity.
- 1.1. Related work: The authors identify limited prior benchmarking against human specialists and report validation of RADnet against senior-radiologist annotations.
2. MATERIALS AND METHODS
The study trains RADnet on annotated brain CT scans using DenseNet with auxiliary hemorrhage segmentation tasks and bidirectional LSTM-based inter-slice modeling. Slice predictions are aggregated into CT-level diagnoses for comparison with radiologist annotations.
- The dataset contains 185 training, 67 validation, and 77 testing brain CT scans obtained from two local hospitals.
- The preprocessing thresholds CT intensities to a brain window and rescales images to isotropic 1 mm × 1 mm resolution without z-axis resampling.
- DenseNet-A adds three auxiliary hemorrhagic-region segmentation tasks to focus classification on relevant features.
- RADnet adds a bidirectional LSTM to DenseNet-A and trains on multiple-slice sequences to model local 3D context while predicting individual slices.
- A CT scan is classified as positive when hemorrhage is predicted in at least three consecutive slices, after which CT-level accuracy, recall, precision, and F1 are calculated.
3. RESULTS
RADnet achieves radiologist-comparable CT-level performance and exceeds two radiologists in recall and F1 score. Its precision is slightly below the minimum radiologist precision.
- 81.82% accuracy, 88.64% recall, 81.25% precision, and 84.78% F1 score were achieved by RADnet at CT level.
- RADnet’s accuracy matched one radiologist, while its recall was higher than two radiologists.
- RADnet showed a higher F1 score than two of the three radiologists.
- 81.25% RADnet precision was slightly lower than the minimum radiologist precision of 82.5%.
4. DISCUSSION & CONCLUSION
RADnet emulates radiologists’ diagnostic approach and achieves comparable accuracy with increased sensitivity, while remaining limited in pathology coverage and real-world validation.
- RADnet emulates radiologists’ method for diagnosing brain hemorrhage from CT scans and is benchmarked against senior radiologists.
- RADnet achieves comparable prediction accuracy to radiologists along with increased sensitivity.
- Segmentation maps provide qualitative indications of hemorrhage severity and focus the model’s attention on relevant classification features.
- The model is unaware of other equally severe brain conditions and should not be interpreted as a replacement for radiologists.
- The method was tested on a limited test set, so its real-world performance requires further experimentation.