Efficient Strategies for Better Imbalance Image Segmentation

Dawood AL CHANTI, Vanessa Gonzalez Duque, Marion CROUZIER, Antoine Nordez, Lilian Lacourpaille, Diana Mateus

Keywords: Semantic Segmentation, Parametric Tversky Loss, Pixel Imbalance, Dilated Filter Re-use

TL;DR: We propose to handle pixel-wise imbalance and to encourage filter reuse in dilation mode to better model the contextual information.

Abstract: We propose a strategy that encourages filter reuse to decrease the total number of learned parameters and to enable training on small dataset efficiently. We also highlight on one of our recent publication (Al Chanti et al., 2021), which handles foreground/background class imbalance by learning adaptively how to penalize False Positives and False Negative pixels, resulting in a faster convergence and better performance. We validate our method on low limb muscle segmentation using volumetric ultrasound.

Paper Type: both

Primary Subject Area: Segmentation

Secondary Subject Area: Validation Study

Paper Status: based on accepted/submitted journal paper

Source Code Url:

https://github.com/DawoodChanti/Efficient-Strategies-for-Better-Medical-Image-Segmentation/blob/main/Filter_Reuse_with_Dilation_and_Parametric_Tversky_Loss.ipynb

https://github.com/DawoodChanti/IFSS-Net

Data Set Url: The current database is still private and in the process of collecting the acknowledgement of the subjects to be made publicly available.

Registration: I acknowledge that publication of this at MIDL and in the proceedings requires at least one of the authors to register and present the work during the conference.

Authorship: I confirm that I am the author of this work and that it has not been submitted to another publication before.