High-Resolution Breast Cancer Screening with Multi-View Deep - - PowerPoint PPT Presentation
High-Resolution Breast Cancer Screening with Multi-View Deep Convolutional Neural Networks Krzysztof J. Geras Joint work with Kyunghyun Cho , Linda Moy , Gene Kim , Stacey Wolfson and Artie Shen . GTC 2017 W HERE DEEP LEARNING IS USEFUL amount
Joint work with Kyunghyun Cho, Linda Moy, Gene Kim, Stacey Wolfson and Artie Shen.
GTC 2017
difficulty of the task amount of data available
difficulty of the task amount of data available
the learning tasks for which deep learning makes a difference
natural image recognition machine translation speech recognition game playing
difficulty of the task amount of data available
the learning tasks for which deep learning makes a difference
natural image recognition machine translation speech recognition game playing
medical image analysis?
R-MLO L-MLO R-CC L-CC
(left cranial caudal) (right cranial caudal) (left mediolateral oblique) (right mediolateral oblique)
We try to mimic predictions of a radiologist.
◮ Class 0: incomplete (≈ 15%). ◮ Class 1: negative (≈ 50%). ◮ Class 2: bening findings (≈ 35%).
We try to mimic predictions of a radiologist.
◮ Class 0: incomplete (≈ 15%). ◮ Class 1: negative (≈ 50%). ◮ Class 2: bening findings (≈ 35%).
Radiologists call these classes BI-RADS (short for Breast Imaging-Reporting and Data System).
You need a lot of data to do deep learning.
You need a lot of data to do deep learning. Publicly available data sets contain about 1k images.
You need a lot of data to do deep learning. Publicly available data sets contain about 1k images. We build our own data set:
◮ 23k exams, ◮ 103k images. ◮ Each image is at least 2600×2000 pixels.
High resolution necessary - computational and engineering challenge.
Multi-view data. How to integrate information?
Classifier p(y|x) Concatenation (256×4 dim) DCN DCN DCN DCN L-CC R-CC L-MLO R-MLO
Classifier p(y|x) Concatenation (256×4 dim) DCN DCN DCN DCN L-CC R-CC L-MLO R-MLO layer kernel size stride #maps repetition global average pooling 256 convolution 3×3 1×1 256 ×3 max pooling 2×2 2×2 128 convolution 3×3 1×1 128 × 3 max pooling 2×2 2×2 128 convolution 3×3 1×1 128 × 3 max pooling 2×2 2×2 64 convolution 3×3 1×1 64 × 2 convolution 3×3 2×2 64 max pooling 3×3 3×3 32 convolution 3×3 2×2 32 input 1
0.0 0.2 0.4 0.6 0.8 1.0 false positive rate 0.0 0.2 0.4 0.6 0.8 1.0 true positive rate BI-RADS 0 BI-RADS 1 BI-RADS 2
0 vs. others: 0.609 1 vs. others: 0.717 2 vs. others: 0.728 Average: 0.685
We can compute the entropy of predictions, H(y|x) = −
p(y′|x) log p(y′|x), and sort examples according to it.
We can compute the entropy of predictions, H(y|x) = −
p(y′|x) log p(y′|x), and sort examples according to it. We will consider the 30% with the lowest entropy to be “confident”.
0.0 0.2 0.4 0.6 0.8 1.0 false positive rate 0.0 0.2 0.4 0.6 0.8 1.0 true positive rate BI-RADS 0 BI-RADS 1 BI-RADS 2
0 vs. others: 0.636 1 vs. others: 0.816 2 vs. others: 0.844 Average: 0.765
1/8 1/4 1/2 1 resolution fraction 0.55 0.60 0.65 0.70 0.75 0.80 AUC average AUC average AUC (confident)
1/10 1/5 1/2 1 data set size fraction 0.50 0.55 0.60 0.65 0.70 0.75 0.80 AUC average AUC average AUC (confident)
We visualise
∂xv
(i,j)
where H(y|x) = −
p(y′|x) log p(y′|x).
◮ We made a first step in the direction of end-to-end breast cancer screening
with neural networks.
◮ We made a first step in the direction of end-to-end breast cancer screening
with neural networks.
◮ It is much harder to learn the “incomplete” (0) class than other classes.
◮ We made a first step in the direction of end-to-end breast cancer screening
with neural networks.
◮ It is much harder to learn the “incomplete” (0) class than other classes. ◮ We need to use the full resolution. A lot more effort is necessary to develop
archiectures appropriate for data of large dimensionality.
◮ We made a first step in the direction of end-to-end breast cancer screening
with neural networks.
◮ It is much harder to learn the “incomplete” (0) class than other classes. ◮ We need to use the full resolution. A lot more effort is necessary to develop
archiectures appropriate for data of large dimensionality.
◮ We need more data.
◮ We made a first step in the direction of end-to-end breast cancer screening
with neural networks.
◮ It is much harder to learn the “incomplete” (0) class than other classes. ◮ We need to use the full resolution. A lot more effort is necessary to develop
archiectures appropriate for data of large dimensionality.
◮ We need more data. (We are currently processing a 10 times bigger data set).
High-Resolution Breast Cancer Screening with Multi-View Deep Convolutional Neural Networks
arXiv:1703.07047