Multi-Class Classification Based On Multi-Loss Strategy and Auxiliary Deep Learning Network with Applications in Medical Imaging
Z Fan*1, S He2, E Chen1,6, S Ruan3, X Wang4, H Li1,5 (1)University of Illinois at Urbana-Champaign, Urbana, IL (2)Washington University in St. Louis, St. Louis, MO (3)University of Rouen, Rouen, France (4)University of Illinois at Chicago, Chicago, IL (5)Carle Foundation Hospital, Urbana, IL (6)Carle Illinois College of Medicine, Urbana, IL
Presentations
SU-E-TRACK 6-5 (Sunday, 7/25/2021) 3:30 PM - 4:30 PM [Eastern Time (GMT-4)]
Purpose: Multi-class classification has wide applications in medical imaging fields. However, the performance decreases due to the issues of inter-class data similarity and data imbalance. Previously proposed deep-leaning methods commonly trained the network through the minimization of single loss function to learn discriminative features for classification, which might increase the risks of overfitting. Therefore, a multi-class classification strategy is proposed to address the issues and improve the classification performance.
Methods: The proposed framework backbone includes a feature-extraction network (Feat-Net) and a multi-class classifier. Two auxiliary networks--reconstruction network (Recon-Net) and discriminator-network (Disc-Net)-- are proposed to support the training of Feat-Net and extract class-differentiable features to improve classification performance. Inspired by the design of GAN, Recon-Net aims to recover the images to be similar to the input images using Feat-Net-extracted features, while Disc-Net aims to distinguish the reconstructed and original images. Three networks are trained iteratively by updating Disc-Net only and updating Feat-Net and Recon-Net jointly through the minimization of pre-defined multiple loss functions. In this way, the Feat-Net can extract more informative features stably from the input images compared to traditional methods. Two case studies, the 3-class COVID-19 classification with X-ray images and the two-class stratification of risks of head-and-neck cancer treatment failures with PET images, were employed to evaluate the framework performance.
Results: The framework achieved superior performances compared to the baseline method with 4.6% and 2.5% improvement on F1-score on two case studies, respectively. Also, it was observed that the effectiveness of the proposed framework depends on the depth of Feat-Net and the size of input images.
Conclusion: This study indicates that the auxiliary networks and multi-loss strategy can improve the performance of multi-class classification. The modularized design allows flexibly tailored architectures fitting for varied applications. Extending the framework to other complex tasks would be investigated in the future.
Handouts
Keywords
Image Analysis, Computer Vision
Taxonomy
IM/TH- Image Analysis (Single Modality or Multi-Modality): Classification methods
Contact Email