This document describes a study using deep learning to detect cerebral microbleeds (CMBs) from susceptibility weighted imaging (SWI). It found that using both SWI and phase images as inputs to a neural network led to better performance than SWI alone. The best model achieved 95.8% sensitivity and 70.9% precision, similar to human experts. This outperformed previous studies and demonstrated the potential of deep learning for medical image analysis tasks.
2. Cerebral Microbleed Detection
• Cerebral microbleed (CMB) detection is important in studying dementia,
stroke and traumatic brain injury (TBI)
• Susceptibility Weighted Imaging (SWI) provides exquisite sensitivity to the
presence of CMB
• Manual detection of CMBs
Time consuming
Prone to errors (false positive or missed CMB)
1
3. • Conventional machine learning algorithms
Require feature engineering (shape, intensity etc.)
Variation in the shape and intensity of CMBs leads to low sensitivity or low specificity
• Current deep learning based algorithms
Utilize SWI (magnitude) images only -> difficulties in separating calcifications and CMBs
• Phase images are valuable for separting CMBs and CMB mimics (e.g. calcifications, veins)
2
Magnitude Phase
: CMB
: Calcification
Cerebral Microbleed Detection
4. CMB Detection using Multi-channel 3D MRI Data
3
1. CMB
Candidate
Detection
using SWI
2. False
Positive
Removal
using Phase
and SWI
: CMB
: False Positive
*
*3D-FRST: 3D Fast Radial Symmetry Transform
5. Why 3D-FRST for CMB Candidate Detection
5
B0
CMBs are close to
spheres on SWI
images, due to the
shape of the induced
field variation.
TE=10ms TE=2.5ms
Magnitude
Phase
6. Data Information and Pre-processing
4
Hemodialysis TBI Stroke Normal Control Total
Training 72 (756)a 80 (522) 2 (2) 0 154 (1280)
Validation 15 (160) 8 (33) 2 (0) 0 25 (193)
Test 13 (95) 9 (14) 9 (59) 10 (0) 41 (168)
Data
Splitting
B0 (T) TE (ms) TR (ms) FA (o) In-plane
Resolution (mm2)
Slice Thickness
(mm)
1.5 or 3 40@1.5T;
20@3T
49 or 50 @ 1.5T;
27 to 34 @ 3T
12 or 15 0.5x0.5 to
0.54x1.07
1.2 to 2.65
Imaging
Parameters
aNumber of Cases (Number of CMBs)
Bias-field
Correction
(for magnitude)
Registration
(Magnitude with
MNI-152 template)
Generate
SWI&QSMb
Interpolation
(to 0.5mm
isotropic)
Intensity
Normalization
(to [-1, 1])
Pre-processing
bQSM: quantitative susceptibility mapping
7. Selection of 3D-FRST Threshold
5
th=0.1
Training and Validation
Selection of the 3D-FRST
threshold using train and
validation data: a trade-off
between sensitivity and FPavg.
Test Data Performance
Sensitivity: 99.4%
FPavg: 276.8 per case
th: 0.09
th=0.09 th=0.08
8. Overall Performance on Test Data
6
The best combination was
phase and SWI (PS model).
The PS model achieved
similar performance to the
most experienced human
rater (Rater3).
Rater1: SWI data processor
Rater2: Radiologist
Rater3: Experienced SWI
data processor (8+ years)
9. Contribution of Individual Channels,
Effects of Model Averaging and Input Scale
7
• QSM > SWI > phase > magnitude
• Simple averaging of multiple models (dashed lines) with the same type of input did not help.
• Smaller input scale reduced the performance.
10. Separating CMBs from Calcification
8
Magnitude Phase SWI QSM
3D-FRST SWI SWI with Candidates SWI Model Prediction Phase+SWI Model Prediction
: calcification
11. Performance of PS Model in Each Case and Error Analysis
9
False positives caused by the asymmetric
prominent cortical veins (APCV) in a stroke case.
Missed lesion due to small size in a hemodialysis case. Missed classified damaged veins in a TBI case.
12. Comparison with Earlier Studies
10
Models Input Number of parameters Sensitivity (%) Precision (%) FP per CMB
Chen et al.1
SWI ~200,000 94.7 72.0 0.4
Dou et al.2
SWI ~1,380,000 93.2 44.3 1.2
SWI Model SWI 71,650 91.7 52.0 0.8
Phase+SWI Model Phase + SWI 72,082 95.8 70.9 0.4
1Chen et al. Toward Automatic Detection of Radiation-Induced Cerebral Microbleeds Using a 3D Deep Residual Network. J Digit Imaging. 2018.
2Dou et al. Automatic Detection of Cerebral Microbleeds From MR Images via 3D Convolutional Neural Networks. IEEE Trans Med Imaging. 2016;35(5):1182–95
13. Discussion
• Limited sample size of different categories
variation in the performance for different types of disease
• Lack of comprehensive comparison with earlier algorithms
used the SWI model as a reference
compared with human raters
• The raw phase images were not available
limited quality of QSM -> the best combination of input channels is phase
and SWI
• Use of conventional 3D-FRST for candidate selection:
fully convolutional network may further improve time-efficiency
11
14. Conclusions
• The use of both SWI and phase images leads to significant improvement in the
model performance.
• Preprocessing is important for utilizing 3D multi-contrast MRI data.
• The best PS model outperformed earlier studies and achieved similar
performance to the most experienced human rater.
• There is great potential of applying deep learning techniques to medical
imaging.
14
15. • A paper based on this study has been published on NeuroImage:
• Liu S, Utriainen D, Chai C, Chen Y, Wang L, Sethi SK, Xia S, Haacke EM. Cerebral
microbleed detection using Susceptibility Weighted Imaging and deep
learning. NeuroImage. 2019 May 20.
• https://doi.org/10.1016/j.neuroimage.2019.05.046
15