project_review

---

01. coordconv solution

---

https://arxiv.org/abs/1807.03247

idea

• coordconv : 좌표정보를 input에 concat 시켜 추가하고 conv 연산을 수행해 모델이 좌표정보 또한 학습하게끔 만듬
  • input에 nomalization한 좌표정보 i,j를 concat

---

01-1. AUTOMATED SEGMENTATION OF PULMONARY LOBES USING Coordination-Guided Deep Neural Networks

---

https://arxiv.org/abs/1904.09106

task & contributions

• propose an automated segmentation of pulmonary lobes using coordination-guided deep neural networks from chest CT images

dataset

• 343 chest CT scans
• size : 256×256×128
• target : five target lobar classes

model

• fully end-to-end 3D deep learning approach
• unet 구조와 유사한데 큰 차이점은 skip connection 사용
• coordconv 적용 : last transition in the decoding path

evaluation

---

02. Focal loss

---

idea

• pt가 클 경우 상대적으로 loss가 pt가 적을 때 보다 크게 감소
• 따라서 상대적으로 잘 분류되지 못한 cls에 집중함

---

02-1. A NOVEL FOCAL TVERSKY LOSS FUNCTION WITH IMPROVED ATTENTION U-NET FOR LESION SEGMENTATION

---

https://arxiv.org/abs/1810.07842

task & contributions

• a novel focal Tversky loss function for highly imbalanced data and small ROI segmentation
• a deeply supervised attention U-Net improved with a multiscaled input image pyramid for better intermediate feature representations.

dataset

• 1.Breast Ultrasound Lesions 2017 dataset B (BUS)

  • 163 ultrasound images of breast lesions from different women
  • average image size is 760 x 570 pixels(resampled to 128 x 128 pixels)
  • a 75-25 train-test split

• 2.ISIC 2018 skin lesion dataset

  • 2,594 RGB images of skin lesion
  • image size of 2166 x 3188 pixels(resampled to 192 x 256 pixels)
  • 75-25 train-test split

focal Tversky loss function (FTL)

• The Tversky index is adapted to a loss function (TL)

* pic : probability that pixel i is of the lesion class c
* pic-: probability pixel i is of the non-lesion class c¯
* same is true for gic and gic¯
*  α,β 는 하이퍼파라미터

• focal Tversky loss function (FTL)

  

• γ : 파라미터[1~3] /best : 4/3

model

• attention gate
• multi scale : avg pooling을 이용해 각 stage의 input으로 추가로 넣어줌
• deep supervision

evaluation