https://github.com/parth-shastri/siim_pneumothorax_segmentation

Unet based model for the SIIM Pneumothorax dataset, (based on the original Unet Architecture).
https://github.com/parth-shastri/siim_pneumothorax_segmentation

Last synced: 15 days ago
JSON representation

Unet based model for the SIIM Pneumothorax dataset, (based on the original Unet Architecture).

• Host: GitHub
• URL: https://github.com/parth-shastri/siim_pneumothorax_segmentation
• Owner: parth-shastri
• Created: 2021-05-29T17:35:13.000Z (over 3 years ago)
• Default Branch: master
• Last Pushed: 2021-06-04T12:22:27.000Z (over 3 years ago)
• Last Synced: 2024-11-11T15:22:55.201Z (2 months ago)
• Language: Python
• Size: 677 KB
• Stars: 0
• Watchers: 1
• Forks: 0
• Open Issues: 0
• Metadata Files:
  • Readme: README.md

Awesome Lists containing this project

README

        
# SIIM_Pneumothorax_segmentation

Unet based model for the SIIM Pneumothorax dataset, (based on the original Unet Architecture). This project is based on the challenge posted on Kaggle.

Statement: Segment the infected chest-Xray images in the dataset. The dataset can be downloaded from https://www.kaggle.com/seesee/siim-train-test.

The data is in DICOM format, with masks in RLE format.

LOSS : Binary Cross-Entropy, pixelwise classification. 

Update : Added Focal loss and also DICE loss and a combo of BCE and DICE Losses.

The Original U-Net architecture is used.

Paper : https://arxiv.org/pdf/1505.04597

TODO : deal with imbalanced dataset

Done - Implemented focal loss to mitigate the data imbalance.

Note : I've trained the task on a subset of images, due to the constraints of device and memory. this will work pretty well on the total data as well.

P.S. added metric from the repository of sidml - https://github.com/sidml/SIIM-ACR-Pneumothorax-Segmentation (my_iou_metric) found it pretty usefull.