IDEC Image Segmentation

Advanced Image Processing Lecture based on Deep Learning in 한국과학기술원 반도체설계교육센터 (KAIST IDEC)

Pneumonia Classification and Pet Data Image Segmentation with U-Net
In 2021 Advanced Image Processing Lecture with KAIST IDEC

DataSet

• Pet Data - 1190 Images (Ragdoll, saint_bernard, scottish_terrier, Siamese, staffordshire_bull_terrier, yorkshire_terrier)
  [University of OXFORD : Open Dataset]
  (https://www.robots.ox.ac.uk/~vgg/data/pets/data/images.tar.gz) (https://www.robots.ox.ac.uk/~vgg/data/pets/data/annotations.tar.gz)

• Pneumonia Data - 1200 Images (Normal, Pneumonia)
  [Identifying Medical Diagnoses and Treatable Diseases by Image-Based : Open Dataset] (http://download.tensorflow.org/data/ChestXRay2017/train/images.tfrec) (http://download.tensorflow.org/data/ChestXRay2017/train/paths.tfrec)

Application

Model	Loss Function	Optimizer	Epoch	Total Loss	Accuracy (Classification) (=F1-Score)	Dataset	Result
U-Net	Sparse Categorical Crossentropy	RMSProp	200	0.0269 (Train)	.	Pet Data	Result_01 (Segmentation)
U-Net	Sparse Categorical Crossentropy	RMSProp	1000	0.0041 (Train)	.	Pet Data	Result_02 (Segmentation)
U-Net Leaky ReLU	Sparse Categorical Crossentropy	RMSProp	1000	0.0048 (Train)	.	Pet Data	Result_03 (Segmentation)
FCN	Binary Crossentropy	Adam	200	1.0335 (Test)	0.8650 (Test)	Pneumonia Data	Result_04 (Classification)
FCN	MSE	Adam	200	0.1464 (Test)	0.8500 (Test)	Pneumonia Data	Result_05 (Classification)
FCN	Binary Crossentropy	Nadam	200	1.6821 (Test)	0.7500 (Test)	Pneumonia Data	Result_06 (Classification)
FCN	Binary Crossentropy	AdaDelta	200	0.3628 (Test)	0.8600 (Test)	Pneumonia Data	Result_06 (Classification)

Result_01

Result_02

Result_03

Sample Dataset

Sample Pet Train Raw Data (each different size)

Siamese_203

yorkshire_terrier_67

Sample Pet Train Scaling Data (160 x 160 px)

Siamese_203

yorkshire_terrier_67

Sample Pneumonia Train Scaling Data (180 x 180 px)

Pneumonia

Reference

https://keras.io/examples/vision/