Page 81 - My FlipBook
P. 81
Brochure 2020
3. Initiated Projects
3.1 Automatic 3D Instance Segmentation and Detection Framework for Cerebral Microbleeds
The ageing population has become a global issue. In particular, dementia places a tremendous burden on
society. Early identi cation of high-risk groups, as well as early diagnosis, can e ectively slow the progression
of dementia through appropriate treatments. The cerebrovascular system is closely linked to the cerebral
nervous system. Cerebrovascular damage can seriously affect functioning of the nervous system and may
even induce neurodegeneration. Therefore, stroke patients are at a high risk of dementia. E ectively detecting
cerebrovascular variations and the location and size of lesions, as well as clarifying the relationships between
cerebrovascular and neurological diseases, could greatly help with diagnoses and treatments of dementia.
Cerebral microbleeding is a cerebrovascular disease that is correlated with loss of cognitive function and
increased risk of stroke (Chung et al., 2016; Chung et al., 2017), and it impacts treatments for ischemic stroke
and the options for subsequent precautionary measures. Cerebral microbleeds can be observed using
magnetic sensitive weighted angiography (SWAN), manifesting in images as low-signal circular or oval-
shaped areas with defined edges (Greenberg et al., 2009). However, other structures may present similar
characteristics in SWAN images, such as normal blood vessels and calci cation, impeding accurate diagnosis.
Moreover, since microbleeds are small and difficult to detect, they are time-consuming to diagnose. The
primary goal of this study is to develop enhanced image processing and deep learning technologies that can
automatically localize and measure microbleeds from SWAN images.
Furthermore, in order to clarify the relationships between cerebrovascular and neurological diseases, we are
studying patients with "somatic dominant cerebral artery vascular disease combined with subcortical cerebral
infarction and white matter lesions (CADASIL)". CADASIL is thought to be genetically inherited. Patients
frequently exhibit multiple microbleeds and high-signal lesions in brain white matter (Chen et al., 2014).
Establishing the relationship between microbleed distribution and white matter lesions in the brain will help
clarify the correlation between cerebrovascular diseases and nervous system function.
We have partnered with clinicians to manually construct a microbleed image database with ground-truth
labeling. A 3D object segmentation model has been developed based on 2D entity segmentation model
technology (Mask R-CNN) (Figure 1) to remedy the insu cient accuracy of other 2D models (such as Retina
U-Net and U-Faster R-CNN). The 3D object segmentation model we have created has an AUC of 0.706 (IoU = 0.5),
which is better than those of existing microbleed detection models (Figure 2). A detailed example is presented
in Figure 3.
79
3. Initiated Projects
3.1 Automatic 3D Instance Segmentation and Detection Framework for Cerebral Microbleeds
The ageing population has become a global issue. In particular, dementia places a tremendous burden on
society. Early identi cation of high-risk groups, as well as early diagnosis, can e ectively slow the progression
of dementia through appropriate treatments. The cerebrovascular system is closely linked to the cerebral
nervous system. Cerebrovascular damage can seriously affect functioning of the nervous system and may
even induce neurodegeneration. Therefore, stroke patients are at a high risk of dementia. E ectively detecting
cerebrovascular variations and the location and size of lesions, as well as clarifying the relationships between
cerebrovascular and neurological diseases, could greatly help with diagnoses and treatments of dementia.
Cerebral microbleeding is a cerebrovascular disease that is correlated with loss of cognitive function and
increased risk of stroke (Chung et al., 2016; Chung et al., 2017), and it impacts treatments for ischemic stroke
and the options for subsequent precautionary measures. Cerebral microbleeds can be observed using
magnetic sensitive weighted angiography (SWAN), manifesting in images as low-signal circular or oval-
shaped areas with defined edges (Greenberg et al., 2009). However, other structures may present similar
characteristics in SWAN images, such as normal blood vessels and calci cation, impeding accurate diagnosis.
Moreover, since microbleeds are small and difficult to detect, they are time-consuming to diagnose. The
primary goal of this study is to develop enhanced image processing and deep learning technologies that can
automatically localize and measure microbleeds from SWAN images.
Furthermore, in order to clarify the relationships between cerebrovascular and neurological diseases, we are
studying patients with "somatic dominant cerebral artery vascular disease combined with subcortical cerebral
infarction and white matter lesions (CADASIL)". CADASIL is thought to be genetically inherited. Patients
frequently exhibit multiple microbleeds and high-signal lesions in brain white matter (Chen et al., 2014).
Establishing the relationship between microbleed distribution and white matter lesions in the brain will help
clarify the correlation between cerebrovascular diseases and nervous system function.
We have partnered with clinicians to manually construct a microbleed image database with ground-truth
labeling. A 3D object segmentation model has been developed based on 2D entity segmentation model
technology (Mask R-CNN) (Figure 1) to remedy the insu cient accuracy of other 2D models (such as Retina
U-Net and U-Faster R-CNN). The 3D object segmentation model we have created has an AUC of 0.706 (IoU = 0.5),
which is better than those of existing microbleed detection models (Figure 2). A detailed example is presented
in Figure 3.
79
