Upgrade to Pro
— share decks privately, control downloads, hide ads and more …
Speaker Deck
Sign up for free
Menu
Search
Features
All features
Private URLs
Password Protection
Custom URLS
Scheduled publishing
Remove Branding
Restrict embedding
Deck Collections
Notes
Features
All features
Private URLs
Password Protection
Custom URLS
Scheduled publishing
Remove Branding
Restrict embedding
Deck Collections
Notes
Explore
Featured decks
Featured speakers
Programming
Technology
Storyboards
Explore
Featured decks
Featured speakers
Programming
Technology
Storyboards
Pricing
Search
Sign in
Sign up for free
Generate Abnor Echo Image
Search
Sponsored
·
Your Podcast. Everywhere. Effortlessly.
Share. Educate. Inspire. Entertain. You do you. We'll handle the rest.
→
Tsung-Yung Lu
September 28, 2018
Research
160
0
Share
Embed
Copy iframe code
Copy JS code
Copy link
Start on current slide
Generate Abnor Echo Image
2018SMIRS內容
Tsung-Yung Lu
September 28, 2018
Transcript
31 slides
生成非正常腎臟超音波影像 提升分類準確性之研究 假體實驗 報告者:義守大學 醫學影像暨放射科學系 盧宗詠 2018/09/29
2 <over view> ➢ 研究動機 < motive> ➢ 目的 <purpose>
➢ 材料與方法 < method> ➢ 結果 <result> ➢ 結論與討論 <conclusion>
3 <研究動機>
4
5 “NORMAL”
6 <目的>
7 <purpose> ➢ 解決影像資料集過少的問題 ➢ 比較人為產生的非正常影像 與實際非正常影像 經由神經網路訓練後的優缺分析 ➢ 探討人為加工的醫學影像是否有助於診斷輔
助系統的開發
8 <材料與方法>
9 < materials > Sonosite 180+ 手提式超音波儀 C60
(5-2MHz) Transducer US-1B ABDFAN腹部超音波假體 造影角度不限,深度為12與15公分之 右腎影像,總計蒐集1400張 (合成影像總計生成1600張)
10 < method> 將影像選取ROI並且正規化 從原始影像資料集隨機取出部分影像並透 過自行編寫的matlab程式加工成非正常影像 建立四組資料集以便於進行訓練與驗證
透過卷積神經網絡建立三種判斷模型 CNN模型具有三個2D卷積層(濾波數分別為32、64、128個, 卷積核大小5*5、3*3、3*3)、三個激活層(其激活函數皆使 用ReLU、三個池化層(池化窗口大小皆為3,3) 藉由驗證資料集比較三種模型的差異 繪製分析圖以方便觀察結果
11
12 < method_01.bmp> 1400
13 < method_02.bmp > 1400 NORMAL X1000 ABNORMAL X400
14 < method_03.bmp> NORMAL X1000 X1600
15 < method_04_train_data.bmp> 1.正常 +非正常 (1100) 2.正常 +合成影像 (2300) 3.正常
+(非正常+合成) (2600) NORMAL X800 ABNORMAL X300 NORMAL X800 ABNORMAL X1500 NORMAL X800 ABNORMAL X1800
16 < method_05_validation_data.bmp> NORMAL 200 ABNORMAL (100+100) X400 validation
17 < method_06_train_CNN.py>
18 < method _07_predict_CNN.py > softmax() “NORMAL” “ABNORMAL”
19 < method_08 _ROC.py > ➢ TP (true positive):分類為非正常,實際上為非正常 ➢
TN (true negative):分類為正常實際上為正常 ➢ FP (false positive):分類為非正常,實際上為正常 ➢ FN (false negative):分類為正常,實際上為非正常
20 < method_09_ROC.py > ➢ 準確度,accuracy: 分類正確的比率 ➢ 敏感度, sensitivity:
非正常被分類成非正常的比率(有病判有病) ➢ 錯誤命中率: 正常被分類成正常的比率(沒病判沒病) ➢ 特異度, specificity : 正常被分類為非正常的比率(沒病判有病) ➢ 陽性預測值 :被分類為非正常,實際上為非正常的比率 ➢ F-measure(F度量,F1) :一種同時兼顧查準率(precision)與查全率 (recall)的度量方式,應用於資訊檢索(information retrieval)領域 的成效評估
21 < method_10_ROC.py> ➢ TPR(敏感度, sensitivity) = TP / P
➢ FPR(錯誤命中率) = FP / N ➢ F-measure(F度量,F1) = (2 x TPR x PPV) / (TPR + PPV)
22 <結果>
23 < result> ➢ F-measure ➢ ROC curve with AUC
(receiver operating characteristic curve)
24 < result02.bmp> nor+abnor nor+syn nor+(abnor+syn) True Positive 116 168
198 True Negative 198 157 158 False Positive 2 43 42 False Negative 84 32 2 accuracy 0.785 0.8125 0.89 sensitivity 0.58 0.84 0.99 specificity 0.99 0.785 0.79 false alarm rate 0.01 0.215 0.21 F-measure 0.7295 0.8175 0.9
25 < result01.bmp> nor+abnor nor+syn nor +(abnor+syn) TPR 0.58 0.84
0.99 FPR 0.01 0.215 0.21
26 < result03_ROC.bmp>
27 <結論與討論>
28 < conclusion > ➢ 生成影像能夠有效的解決非正常影像資料 量不足的問題 ➢ 生成影像作為訓練資料可以提升分類模型 的準確性
29 < discussion > ➢ 使用更為複雜的神經網路或者不同的影像生成 方式是否能使準確度提升? ➢ 若生成影像由專業的臨床醫師設計是否能使準 確度更高?
➢ 是否有更好的辦法能夠有效的提升預測模型的 準確度?
THANKS FOR LISTENING github.com/higumalu
[email protected]
Q & A
More Decks by Tsung-Yung Lu
See All by Tsung-Yung Lu
DICOM RT Dose
higumalu
0
110
The GEMPix detector
higumalu
0
67
Respiratory Gating for Radiotherapy
higumalu
0
740
Cholescintigraphy
higumalu
0
230
DL in MRI
higumalu
0
120
Cardiac CT
higumalu
0
330
Class Report of PETCT Model
higumalu
0
120
淺談影像處理
higumalu
0
180
Tc99m
higumalu
0
190
Other Decks in Research
See All in Research
[最先端NLP勉強会2026] Checklists Are Better Than Reward Models For Aligning Language Models
nzw0301
1
350
J-STAGEの現況と全文XML登載必須化について
xspa2012
0
410
Prithvi-EO-2.0
satai
3
120
[SNLP2026] Scaling up Test-Time Compute with Latent Reasoning: A Recurrent Depth Approach
wataruuuuu
0
370
The Conductor Pattern -Multi-Granularity Feedback for Creative Agents-
shibuiwilliam
2
120
実例から見るLLMのマンガ理解:実務VQAタスクによる長期的文脈と視覚情報の定性評価
kzmssk
0
230
Volatilité des marchés et sécheresse historique, l’élevage bovin viande sous contraintes en 2026
institutdelelevage
PRO
0
240
L’albédo:levier d’atténuation du changement climatique dans les systèmes fourragers
institutdelelevage
PRO
0
170
CDCL を用いた MILP の厳密解法
imai448
0
330
HackSick vol.7 LT資料【LLMアーキテクチャ入門・事前学習時の躓き所解説】 スパースなAttention・状態空間モデル
rikkabotan7
0
240
Santé et fertilité des sols : ce que nous disent les pratiques dans les systèmes de polyculture-élevage
institutdelelevage
PRO
0
160
AI時代に学術情報をどう探し、どう使うか: 研究の様々な場面と一研究者の経験から
masao
0
210
Featured
See All Featured
Chasing Engaging Ingredients in Design
codingconduct
0
400
Scaling GitHub
holman
464
140k
We Have a Design System, Now What?
morganepeng
55
8.4k
[Rails World 2026] Durable orchestration on Rails: from continuation to workflow
palkan
4
1.3k
SEO in 2025: How to Prepare for the Future of Search
ipullrank
3
3.9k
The State of eCommerce SEO: How to Win in Today's Products SERPs - #SEOweek
aleyda
2
12k
Building AI with AI
inesmontani
PRO
1
1.4k
How to audit for AI Accessibility on your Front & Back End
davetheseo
0
640
AI Search: Implications for SEO and How to Move Forward - #ShenzhenSEOConference
aleyda
1
1.4k
Game over? The fight for quality and originality in the time of robots
wayneb77
1
330
Heart Work Chapter 1 - Part 1
lfama
PRO
11
37k
Mobile First: as difficult as doing things right
swwweet
225
10k