A Data-Adaptive Loss Function for Incomplete Data and Incremental Learning in Semantic Image Segmentation.

Area of Science:

• Medical image analysis
• Deep learning in healthcare
• Radiation therapy applications

Background:

• Deep learning, particularly convolutional neural networks, has advanced medical image analysis.
• Collecting large, annotated medical datasets for training is costly and time-consuming.
• Existing models struggle with evolving requirements, such as incorporating new anatomical structures in radiation therapy.

Purpose of the Study:

• To propose a novel loss function addressing imbalanced datasets, partially labeled data, and incremental learning in medical imaging.
• To enable deep learning models to adapt to new structures without complete retraining.
• To reduce the computational and logistical burden of developing and maintaining medical image analysis systems.

Main Methods:

• Development of a new loss function designed to utilize all available data, including partially annotated samples.
• Implementation of an incremental learning strategy allowing models to incorporate new anatomical structures.
• Experimental validation on a large in-house dataset.

Main Results:

• The proposed loss function effectively handles imbalanced and partially labeled medical image datasets.
• The method demonstrates successful adaptation in an incremental learning scenario, incorporating new organ delineations.
• Performance is comparable to baseline models, with significant reductions in training time and maintenance overhead.

Conclusions:

• The novel loss function offers an efficient solution for training deep learning models in medical imaging, especially within dynamic fields like radiation therapy.
• This approach facilitates the continuous improvement of decision support systems by enabling seamless integration of new data and structures.
• The method reduces the practical challenges associated with data acquisition and model management in medical AI.