.png){kind=icon}
Return
A dynamic loss function for improved semantic segmentation
.png){kind=icon}
PRE
AI
DOI：10.1007/s00500-023-08818-1.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
The effectiveness of the training process significantly impacts the performance of a machine learning (ML) model. The loss function plays an important role during training as it determines the learning curve of the ML model and offers recommendations for enhancing the model's optimization. The choice of the loss function is crucial for models employed in tasks, such as image classification and semantic segmentation. In the process of segmentation, data are extracted from different image features, which can lead to a diminished gradient value that hampers model training, particularly in high-dimensional scenarios using conventional loss functions. Moreover, as the number of classes increases, major objects often overlap minor objects, leading to false and imprecise detection. To address these issues, a dynamic loss function is needed that can self-adjust to handle class imbalance problems. This paper introduces a dynamic loss function that can adjust itself throughout the model's training process, aiming to tackle the issue of class imbalance. Experiments were conducted on three imbalanced medical image datasets to compare the performance of the proposed dynamic loss function with state-of-the-art loss functions using different ML models. The results demonstrate that during the training process proposed dynamic loss function manages to significantly reduce the convergence time of the model by sharply minimizing the loss. Moreover, during validation process these ML models showed improved performance in terms of accuracy, recall and precision.
Keywords:
Loss function
Hyperparameter
Semantic segmentation
Journal
IF:
2.5
Papers:
1.0W
Citations:
2.1W
Organization
Cited Papers
ShareSave
Electrochemically assisted micro localized grafting of aptamers in a microchannel engraved in fluorinated thermoplastic polymer Dyneon THV.png)
RSC Advances
RSC AdvancesIF0
PREAI
ShareSaveAnatomyNet: Deep learning for fast and fully automated whole-volume segmentation of head and neck anatomyMEDICAL PHYSICS
MEDICAL PHYSICSIF3.2
PREAI
ShareSaveShareSave
View more