Aims and objectives
Artificial intelligence and deep learning techniques are able to detect and localize lesions in head CT scans.
The localization output of the algorithms is usually at pixel level.
radiologists usually report locations of lesions in terms of anatomical regions e.g,
intracranial hemorrhage in left temporal region. In this study,
we aim to create an algorithm which characterizes the affected anatomical regions of a intracranial hemorrhage identified by such an algorithm head,...
Methods and materials
The hemorrhage segmentation algorithm produces pixel mask for localization of intracranial hemorrhage.
To determine the corresponding affected cerebral/cerebellar regions,
the algorithm needs to be aware of anatomy of brain.
To this end,
we created anatomical atlases for five anonymized head CT scans.
Each of theese atlases were marked with marked with the following anatomical regions: Left/right frontal region Left/right temporal region Left/right parietal region Left/right occipital region...
We annotated anatomical atlases for another two head CT scans to evaluate accuracy of the multi-atlas segmentation algorithm.
The average dice score of 0.58 was obtained across different anatomical regions. We evaluated the predicted anatomy locations by manually comparing them to the locations reported by radiologists on 30 head CT scans with intracranial hemorrhage.
We have found that the locations match for 25 scans resulting in an accuracy of 86.20%.
In this work,
we developed algorithms to predict anatomical location given a head CT scan and determine affected lobes/regions with a hemorrhage.
"Deep learning algorithms for detection of critical findings in head CT scans: a retrospective study." The Lancet 392.10162 (2018): 2388-2396. Grewal,
"RADnet: Radiologist level accuracy using deep learning for hemorrhage detection in CT scans." Biomedical Imaging (ISBI 2018),
2018 IEEE 15th International Symposium on .