Iterative confidence relabeling with deep ConvNets for organ segmentation with partial labels.

Area of Science:

• Medical Image Analysis
• Deep Learning
• Computer Vision

Background:

• Deep Convolutional Neural Networks (ConvNets) require extensive labeled datasets for training.
• Pixel-level annotation of medical images is costly, time-consuming, and requires specialized expertise.
• Existing segmentation masks often cover only specific anatomical structures, limiting their utility.

Purpose of the Study:

• To develop a method for training deep learning models on partially labeled medical image datasets.
• To address the challenge of expensive and limited pixel-level annotations in medical imaging.
• To improve the efficiency and applicability of medical image segmentation techniques.

Main Methods:

• Proposed a strategy to identify correct pixels and train Fully Convolutional Neural Networks (FCNs) with a multi-label loss.
• Introduced an iterative confidence self-training approach for relabeling missing pixel labels.
• Developed a confidence network to measure uncertainty and guide the relabeling process (INERRANT).

Main Results:

• INERRANT demonstrated robust performance on partially labeled datasets, comparable to models trained on fully labeled data.
• The method showed effectiveness even with large proportions of missing labels across multiple datasets.
• Highlighted the significance of the iterative learning scheme and confidence measure for optimal results.

Conclusions:

• INERRANT offers an effective solution for leveraging partially labeled medical image data in deep learning.
• The approach significantly reduces the need for extensive manual annotation, making segmentation more accessible.
• Demonstrated a practical application by enriching limited fully labeled data with publicly available partially labeled data.