Intracranial Hemorrhage Detection and Sub-Type Classification Using a Dual-Branch CEPTION–KNNFramework
No Thumbnail Available
Date
2025-07
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Addis Ababa University
Abstract
Intracranial hemorrhage (ICH) is any bleeding inside the skull. It is a life-threatening
disease that requires timely diagnosis and urgent medical treatment. Most often, diagnosis
of head CT is common. However, there is a problem of delayed, missed diagnosis
and misdiagnosis due to subtle, highly variable size, shape, and location of ICH and subtypes.
Workload also leads to interpretation errors which are misdiagnosis and missed
diagnosis. This research leverages a two-branch convolutional neural network (CNN)
architecture—each branch based on Xception—and then fuses their learned features via
concatenation before a final K-Nearest Neighbors (KNN) classification step. Such a hybrid
approach combines the power of deep feature extraction (CNN) with the flexibility
and interpretability of traditional machine learning (KNN). Using an extremely large
data set that encompasses approximately 4.5 million slices, this deep learning algorithm
passed three experiments. The first experiment is Normal vs. Abnormal Classification,
which is a binary classification task where each CT slice is labeled as either ’normal’
or ’abnormal.’ This experiment successfully distinguishes between normal and abnormal
head CT scans with an accuracy of approximately 0.94, with an F1 score near 0.947. The
second experiment, multiclass Classification of distinct hemorrhage types, proves largely
successful. The model excels in detecting intraparenchymal, subdural, subarachnoid, and
intraventricular bleeds, as well as normal scans, each class achieving an F1-score above
0.82. Normal scans also see a robust F1 (0.934). Class-specific accuracies remain high
(roughly 0.95–0.97 for most). However, accuracy alone can be misleading in imbalanced
scenarios or if some hemorrhage types are comparatively rare. Classes like intraparenchymal,
intraventricular, subarachnoid, and subdural show high precision (0.81) and recall
(0.80), reflecting consistent identification of these bleeds. The third and final experiment
assessed how well the end-to-end pipeline generalizes to completely unseen data collected
from Wudassie Diagnostic Center and Yanet General Hospital, Addis Ababa, Ethiopia.
Of the 15 cases at Wudassie Diagnostic Center, 8 were correctly classified, yielding an
overall accuracy of 0.53 and of the 23 cases at Yanet General Hospital, 16 were correctly
classified, yielding an overall accuracy of 0.65. These figures align closely with the
confusion-matrix analysis and underline the model’s promise for triage support in local
hospitals. The findings hold immediate relevance for emergency triage in health-care
systems where certified neuroradiologists are scarce. This tool reliably highlights acute
intracranial hemorrhage with sensitivity above 0.90. In addition, the system’s slicewise
probability maps and subject-level majority decisions augment radiologists rather
than replacing their judgment, thus reducing cognitive load without affecting professional
agency.
Description
Keywords
words:Intracranial hemorrhage (ICH), Deep Learning, Head CT, Image classification