Intracranial Hemorrhage Detection in CT Scans using Deep Learning
Problem
Patient mortality in intracranial hemorrhage depends on prompt diagnosis from radiologist assessment of CT scans; a deep learning model was built to accelerate identification.
Why it matters
Faster, accurate hemorrhage detection could reduce patient mortality from delayed diagnosis.
What the AI did
Trained a ResNet-50-based convolutional neural network on 752,803 DICOM files collected by RSNA to predict hemorrhage type in CT scans.
Results
93.3% multiclass accuracy; 76% average per-class recall; 86% average recall at 70% precision via threshold tuning
Impact
Demonstrates feasibility of deep learning for hemorrhage classification on a large dataset; no clinical deployment stated.
Record
| Venue | IEEE International Conference on Big Data Computing Service and Applications (BigDataService) |
|---|---|
| Published | 2020 |
| Work type | conference-paper |
| Authors | Tomasz Lewick, Meera Kumar, Raymond Hong, Wencen Wu |
| Citations | 36 (OpenAlex, at import time) |
| References | 14 |
| Model or method | ResNet-50 (convolutional neural network) |
| Organization | Radiological Society of North America (dataset source) |
| Primary source | https://doi.org/10.1109/bigdataservice49289.2020.00033 |
| Open access | closed |
| Keywords | intracranial hemorrhage, CT scans, deep learning, ResNet-50, medical imaging |
Importance rating
significant — thousands to millions affected
Abstract states patient mortality in intracranial hemorrhage depends on prompt CT-based diagnosis; model trained on 752,803 DICOM files with 93.3% accuracy, a significant clinical need.