Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Interactive Image Selection and Training for Brain Tumor Segmentation Network

Matheus A. Cerqueira, Flávia Sprenger, Bernardo C. A. Teixeira, Alexandre X. Falcão

TL;DR

The paper tackles the challenge of brain tumor segmentation under limited annotated data by introducing an interactive FLIM-based workflow to train a lightweight encoder within a U-shaped network. An expert marks regions in selected images to learn convolutional filters directly, and subsequent images are chosen based on how poorly their regions activate with the current filters, iterating up to a fixed limit. The encoder is learned with FLIM while the decoder is trained by backpropagation, and the approach achieves competitive Dice scores compared to training on the full dataset and to golden-standard models, using a fraction of the labeled data. This method reduces labeling burden and offers a practical pathway toward efficient, clinically applicable BTS models.

Abstract

Medical image segmentation is a relevant problem, with deep learning being an exponent. However, the necessity of a high volume of fully annotated images for training massive models can be a problem, especially for applications whose images present a great diversity, such as brain tumors, which can occur in different sizes and shapes. In contrast, a recent methodology, Feature Learning from Image Markers (FLIM), has involved an expert in the learning loop, producing small networks that require few images to train the convolutional layers. In this work, We employ an interactive method for image selection and training based on FLIM, exploring the user's knowledge. The results demonstrated that with our methodology, we could choose a small set of images to train the encoder of a U-shaped network, obtaining performance equal to manual selection and even surpassing the same U-shaped network trained with backpropagation and all training images.

Interactive Image Selection and Training for Brain Tumor Segmentation Network

TL;DR

The paper tackles the challenge of brain tumor segmentation under limited annotated data by introducing an interactive FLIM-based workflow to train a lightweight encoder within a U-shaped network. An expert marks regions in selected images to learn convolutional filters directly, and subsequent images are chosen based on how poorly their regions activate with the current filters, iterating up to a fixed limit. The encoder is learned with FLIM while the decoder is trained by backpropagation, and the approach achieves competitive Dice scores compared to training on the full dataset and to golden-standard models, using a fraction of the labeled data. This method reduces labeling burden and offers a practical pathway toward efficient, clinically applicable BTS models.

Abstract

Medical image segmentation is a relevant problem, with deep learning being an exponent. However, the necessity of a high volume of fully annotated images for training massive models can be a problem, especially for applications whose images present a great diversity, such as brain tumors, which can occur in different sizes and shapes. In contrast, a recent methodology, Feature Learning from Image Markers (FLIM), has involved an expert in the learning loop, producing small networks that require few images to train the convolutional layers. In this work, We employ an interactive method for image selection and training based on FLIM, exploring the user's knowledge. The results demonstrated that with our methodology, we could choose a small set of images to train the encoder of a U-shaped network, obtaining performance equal to manual selection and even surpassing the same U-shaped network trained with backpropagation and all training images.
Paper Structure (13 sections, 4 figures, 3 tables)

This paper contains 13 sections, 4 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Related Works
  3. Image Selection
  4. Feature Learning from Image Markers
  5. Method
  6. Experiments
  7. Datasets
  8. Adopted Architecture
  9. Encoder and Decoder training
  10. Evaluation Metrics
  11. Golden Standard Models
  12. Results
  13. Conclusion

Figures (4)

  • Figure 1: Our interactive methodology. Learn filters from image markers, then select the next image based on the performance of the learned filters.
  • Figure 2: Image selection criteria and examples of good and bad activation superimposed on the original image.
  • Figure 3: Model performance based on the number of encoder's training images.
  • Figure 4: Example of sample and binary image from their best feature:(a) first image, second image with a bad (b) and good feature (c).