75th Annual Conference of the
Indian Radiological and Imaging Association
IRIA 2023, Amritsar
TITLE: Deep Learning Model for localization of Intracranial Hemorrhages: A step towards effective
pre-read result generation to assist the Radiologist
AUTHOR : Dr. Pallavi Rao
EMAIL: palloo.m@gmail.com, pallavi.rao@imagecorelab.com
CO-AUTHORS: Dr. Arjun Kalyanpur, Dr Anjali Agrawal, Mr Ishan Shah, Mr. Vanchhit Kumar Dubey,
Mr Raqeeb Ansari
INSTITUTION:Image core lab, Teleradiology Solutions, Teleradtech, Bangalore
75th Annual Conference of the
Indian Radiological and Imaging Association
IRIA 2023, Amritsar
Background:
There are several DL algorithms designed to detect intracranial hemorrhages (ICH). While basic detection is done
efficiently, precise localization of hemorrhage akin to a human radiologist remains a challenge.
Learning objectives:
The objective of this study is to design and assess efficiency of a deep learning (DL) model to provide exact location of
intracranial hemorrhages on non-contrast CT (NCCT) brain studies.
Procedure and methods:
DL based classification-segmentation approaches were used to classify types of ICH and provide anatomical locations.
We used 266 NCCT head studies with ICH annotations performed by radiologists with greater than 10 years of
experience.
The brain window was used with pre-processing such as windowing and hue adjustment.
Training and validation studies were divided in the ratio of 80:20, i.e., 213 cases for training and 53 cases for validation.
Testing was carried out in separate 291 cases.
Mobile-Net and U-Net architectures were used for classification and segmentation respectively.
The detection was done as a two-step process, the first being ICH classification, followed by a sub-classification for their
anatomical locations.
A different patch overlay based pre-processing step using a segmentation model was used for sub-classification.
75th Annual Conference of the
Indian Radiological and Imaging Association
IRIA 2023, Amritsar
Mobile-net: Used for
classification of hemorrhage
• Mobile-net model, which has been used for classification, is used with pre-trained weight on ‘imagenet’ dataset. We further added
dropout regularizer to avoid overfitting followed by classification layers with sigmoid activation.
• 5 sigmoid activation functions were used for classification corresponding to each class- Intraparenchymal hemorrhage(IPH), Subdural
hemorrhage (SDH), Intraventricular hemorrhage ( IVH), Subarachnoid hemorrhage (SAH) and Epidural hemorrhage (EDH).
• Image segmentation is a commonly used technique in digital image processing and analysis to partition an image into multiple parts or
regions, often based on the characteristics of the pixels in the image.
75th Annual Conference of the
Indian Radiological and Imaging Association
IRIA 2023, Amritsar
● UNet model which is a fully convolutional network model specifically developed for biomedical image
segmentation, is used for binary segmentation. The model was trained on image size 512x512.
● It's an encoder-decoder type model which can provide fast and precise segmentation of images.
Unet: Used for segmentation
of hemorrhage
WORKFLOW
75th Annual Conference of the
Indian Radiological and Imaging Association
IRIA 2023, Amritsar
Case reports
75th Annual Conference of the
Indian Radiological and Imaging Association
IRIA 2023, Amritsar
Case reports
75th Annual Conference of the
Indian Radiological and Imaging Association
IRIA 2023, Amritsar
Findings and Results
Overall result
291 positive studies of ICH on NCCT were used for the validation of
our model and reviewed by radiologists. A score of 0 to 4 for
acceptability of DL output was given for each study. The validation
result found 48.89% of cases achieved a score of 4 which states
complete agreement, 12.14% of cases with a score of 0 which suggests
complete disagreement. 22.74%, 10.42% and 7.1% of cases were
given partial agreement scores of 3, 2 and 1 respectively. Overall,
87.86% of studies had some level of correct predictions.
Agree Disagree Partially Agree
39%
48%
13%
The chart represents the comparative
number of agreement, disagreement and
partial agreement for each type of ICH
75th Annual Conference of the
Indian Radiological and Imaging Association
IRIA 2023, Amritsar
• This radar chart shows the region of complete
agreement, partial agreement and disagreement. The
poles of the region are IPH, SDH, IVH, SAH and EDH
and demonstrate the combination that DL model is most
likely to predict accurately and what would likely be
missed.
• If the class and location predicted have a combination
of SDH and EDH, then the probability of complete
agreement is high (SDH has complete agreement in 47
cases, partial agreement in 32 cases out of 90 cases). In
EDH similar behaviour is seen, complete agreements
were higher than partial agreement being 13 and 12 out
of 32 cases, respectively.
• The same trend of complete agreement being greater
than partial agreement is seen in IVH with SAH.
• The combinations involving IPH with other
hemorrhages have a higher incidence of partial
agreement.
The number lines of pentagon represents the
number of cases belonging to particular class or the
combinations.
75th Annual Conference of the
Indian Radiological and Imaging Association
IRIA 2023, Amritsar
● Pitfalls:
Current version of the algorithm limits prediction to two types and locations of
hemorrhage per study. Higher number of hemorrhages may be omitted in the DL
output. We intend to overcome this and further improve accuracy in the next
version.
● Conclusion:
We conclude that accurate DL classification and localization of ICH can be
effectively performed to assist radiologists and help prioritise positive cases and
reduce turnaround time.