Unsupervised Deep Learning Applied to Breast Density Segmentation and Mammographic Risk Scoring

Journal Review
Unsupervised Deep Learning Applied to Breast Density Segmentation
and Mammographic Risk Scoring
Jinseob Kim
December 27, 2017
Jinseob Kim Journal Review December 27, 2017 1 / 43

1 Introduction
2 Method
3 Result
4 Discussion

Introduction
Introduction

Introduction
The goal of this paper
Automatically learns features for images, which in our case are
mammograms
CSAE: convolutional sparse autoencoder
Sparse autoencoder within a convolutional architecture
Unlabelled data로 추상적인 feature학습 + label정보 이용해 분류: 2
optimization process
많은 unlabbeled data 이용 가능.
한번에 다 하는 것보다 Fast and stable

Introduction
2 tasks
1 The automated segmentation of percentage mammographic
density (PMD)
2 Characterize mammographic textural (MT) patterns with the goal
of predicting whether a woman will develop breast cancer.
Structural information of breast tissue
Heterogeneity > density.
Manual vs automated
Harder than MD scoring,
The label of interest (healthy vs. diseased) is deﬁned per image and not
per pixel (e.g., fatty vs. dense)

Introduction
Model 요약
1 Multiscale denoising autoencoders
2 Convolutional architecture
3 Novel sparsity term to control the model capacity

Introduction
Figure 1: Multiscale

Method
Method

Method
3 part
1 Generating input data
Multiscale
2 Model representation
CNN
3 Parameter learning
Sparse autoencoder: novel sparsity regularizer

Method
Problem
1 Entire image를 input으로 쓰긴 어렵다..computational burden
2 Downsampling은 안된다. Fine scale로만 확인할 수 있는 feature..
Learn a compact representation for local neighbors (or patches) from
the image

Method
Figure 2: Patch

Method
Patch creation
1 Image resize: 50 pixel/mm
2 Sampling 48000 patches from whole data.
Restricted to 24pixel × 24pixel size
3 Density scoring
10% from the background and the pectoral muscle
45% from the fatty breast tissue
45% from the dense breast tissue.
4 Texture scoring
50% from the breast tissue of controls
50% from the breast tissue of cancer cases.

Method
m = 24, M = 1, 4 layer: Conv + Maxpool + Conv + Conv
c = 1 (흑백): 칼라면 3
If t = 1, 인접한 m × m. t = 4, 더 큰 smoothed 그림에서 every 8th
pixel 선택.
K = {50, (50), 50, 100}, kernal size= {7, 2, 5, 5}

Method
Multiscale input data
Capture long range interactions
Gaussian scale space
I(u; σt) = [I ∗ Gσt ](u)
Gσt =
1
2πσt
e
−(x2+y2)
σt
σt =
t−1
i=0
δ2i
δ : downsampling factor=2

Method

Method
3 step from z(l)
to z(l+1)
1 Extract sub-patches (called local receptive fields)
2 Feature learning: Learn transformation parameters (or features) by
autoencoding the local receptive fields
3 Feature encoding: Transform all local receptive fields using the learned
features from step 2.
Last: Softmax classifier(multinomial logistic regression)

Method
Sparse autoencoder
전체 architecture에서 training할수도 있지만..
Use unsupervised learning: autoencoder
Sparse
Enables to learn a sparse overcomplete representation
Input보다 size가 큰 feature

Method
Standard autoencoder
https://wikidocs.net/3413

Method
Tied weight autoencoder
Input(r):Subpatch(d × d of c channel)
Encoder: K개 feature로 표현 (K < cd2)
a ≡ g(r) = φ(Wr + b)
φ(x) = max(0, x)
Decoder: K개 feature를 이용해서 다시 Input을 표현.
f (a) = ψ(W a + ˜b)
Tied weight

Method
Learned features as input of next layer
1 subpatch 당 K차원 feature
1 patch 당 (m − d + 1) × (m − d + 1) × K 차원의 feature
Next layer의 input

Method
Estimation

Method
Sparse overcomplete representations
N(Basis Vectors) > dimensions of the input
K > cd2
This study, K = {50, (50), 50, 100}, kernal size= {7, 2, 5, 5}
http://mlsp.cs.cmu.edu/courses/fall2013/lectures/slides/class15.
sparseovercomplete.pdf

Method
Type of Sparsity
1 Population sparsity: 적은 갯수의 feature로 input 설명
Compact encoding per example
2 Lifetime sparsity: 한 feature가 여러 input을 설명하는데 쓰이지 않음.
Example-speciﬁc features

Method
Estimation: Novel sparsity term

Method

Method
Experiments and Datasets
2 diﬀerent tasks (MD, MT)
First segmented mammograms into background, pectoral muscle,
and breast tissue region
3 diﬀerent mammographic datasets
Density Dataset(493 mammograms of healthy women): radiologist
Texture Dataset(MMHS): trained observer
Dutch Breast Cancer Screening Dataset: Software(Volpara)

Result
Result

Result
MD: Density datasets
Initial output: 해당 pixel(patch)이 dense tissue class일 확률
빠른 training을 위해 dense tissue data를 많이 포함: overestimation
가능성
dense 판단 기준 threshold를 0.5에서 0.75로 올림

Result
Image-wise average of the Dice coeﬃcients
2 · |A ∩ B|
|A| + |B|
A: automated, B: radiologisat

Result

Result
MD: Dutch Breast Cancer Screening Dataset
Density dataset으로 얻은 모형을 이 dataset에 적용

Result
MT: Texture Dataset
Initial output: 해당 pixel(patch)이 cancer class일 확률.
MT score per image: Breast area에서 patch 500개 랜덤으로 뽑아
Score 구한 후 평균.
랜덤으로 해도 AUC차이 0.01미만임을 확인.

Result
MT: Dutch Breast Cancer Screening Dataset

Discussion
Discussion

Discussion
Summary
Present an unsupervised feature learning method for breast density
segmentation and automatic texture scoring.
Can learn useful features
After adapting a small set of hyperparameters (feature scales, output
size, and label classes), the CSAE model achieved state-of-the-art
results on each of the tasks

Discussion
CSAE vs Classical CNN
The usage of unsupervised pre-training
성능증가

Discussion
Limitation: MT scoring
하나의 mammogram은 어떤 위치든 다 같은 label: Case vs Control
Assumed that texture changes are systemic and occur at many
locations in the tissue

Discussion
Conclusion: Future idea
Image의 여러 부분 정보(patch)를 합쳐서 하나의 label로 매핑.
더 많은 데이터, 컴퓨팅 성능 필요
Easily adjusted to support 3D data

Unsupervised Deep Learning Applied to Breast Density Segmentation and Mammographic Risk Scoring

More Related Content

What's hot (7)

Similar to Unsupervised Deep Learning Applied to Breast Density Segmentation and Mammographic Risk Scoring (20)

More from Jinseob Kim (20)

Recently uploaded (20)

Unsupervised Deep Learning Applied to Breast Density Segmentation and Mammographic Risk Scoring