Multi-Task Deep Learning for Image Segmentation Using Recursive Approximation Tasks.

Area of Science:

• Computer Vision
• Machine Learning
• Medical Imaging

Background:

• Fully supervised deep neural networks for image segmentation demand extensive pixel-level annotations, which are costly and time-consuming to generate.
• Existing methods often struggle with the annotation bottleneck in acquiring precise segmentation masks.

Purpose of the Study:

• To develop a multi-task learning method that significantly reduces the requirement for pixel-level labels in image segmentation.
• To propose a novel framework that treats segmentation as recursively defined approximation subproblems.

Main Methods:

• A multi-task learning framework incorporating a segmentation task, a recursive approximation task, and auxiliary tasks trained with sparse annotations.
• The recursive approximation task iteratively refines partial object masks towards accurate boundaries using learned statistics and segmentation task guidance.
• Training utilizes a small set of precisely segmented images and a large set of coarsely labeled images.

Main Results:

• Demonstrated efficiency in three distinct applications involving microscopy and ultrasound images.
• Achieved effective image segmentation with significantly reduced annotation effort.
• The proposed method successfully learns from partial object masks and coarse labels.

Conclusions:

• The developed multi-task learning approach offers an efficient and cost-effective solution for image segmentation.
• This method alleviates the annotation burden, making deep learning models more accessible for segmentation tasks.
• The framework shows promise for applications requiring segmentation of microscopy and ultrasound imagery.