/Multi-Object-Semantic-Segmentation

Multi Object Semantic Segmentation of Lung Dataset (2 Classes: Left or Right)

Primary LanguageJupyter Notebook

Multi-Object-Semantic-Segmentation

Task: Segment Left and Right Lung from Lung Dataset

Presentation link: https://docs.google.com/presentation/d/1gTv_T3emhZpL3ZO1yKkmYnJcfkhZl6ke1m0ANZmFM7g/edit#slide=id.g13c4dc079d2_0_356

Data

img_mask

(Contact me via e-mail if you need segmented data)

Pre-processing Images

Models

  • PSPNet (Pyramid Scene Parsing Network)
  • ResNet101

FINAL MODEL

image

  • Model: PSPNet (Pyramid Scene Parsing Network)
  • Pretrained weights on: ADE20k
  • Optimizer: Adam
  • LR scheduler: Cosine Annealing
  • Criterion : 0.01xSoft Dice Loss + 0.99xBoundary Loss + 0.3xAuxiliary Loss
  • Performance: 0.98 dice on test set

Results

Model (epoch) Augmentation Average Dice Coefficient (on Test score) model link
PSPNet (epoch: 50) Random Mirror (num. of train imgs: 794) mean dice: 0.9425 (left lung 0.9538, right lung: 0.9312)
PSPNet (epoch: 25) Cutmix (num. of train imgs: 367) mean dice: 0.9501 (left lung: 0.9622, right lung: 0.9380)
PSPNet (epoch: 30) Albumentation1 (num. of train imgs: 367) mean dice: 0.9473 (left lung:0.9601 right lung: 0.9345)
PSPNet (epoch: 25) Cutout (num. of train imgs: 367) mean dice: 0.4597 (left lung: 0.2758, right lung: 0.6436)
PSPNet pretrained on cutmix (25) (epoch: 10) + HD losss None mean dice: 0.613 (left lung: 0.5941 right lung: 0.6328)
PSPNet pretrained on cutmix (25) (epoch: 10) + minimized HD losss + increased weight decay + decreased lr None mean dice: 0.9554 (left lung: 0.9645 right lung: 0.9464 )
PSPNet pretrained on cutmix (25) (epoch: 25) + minimized HD losss + increased weight decay + decreased lr None mean dice: 0.9555 (left lung 0.9645 right lung: 0.9464 )
PSPNet pretrained on cutmix (25) (epoch: 5) + minimized BD losss + trainable weights on loss Histogram equalizer mean dice: 0.9612 (left lung 0.9671 right lung: 0.9553 ) link
PSPNet pretrained on cutmix (25) (epoch: 10) + weighted DC + BD Histogram equalizer mean dice: 0.9498 (left lung 0.9560 right lung: 0.9436 )
PSPNet Cosine Annealing (epoch: 50) + DC + BD Histogram equalizer mean dice: 0.98 (left lung: 0.98141 right lung: 0.97686) link

Contact

If you need codes or have any issue, please contact me via e-mail (superbunny38 at gmail dot come) with your needs.