0%

(2017) Boosted cascaded convnets for multilabel classification of thoracic diseases in chest radiographs

Posted on In , ,

Keyword [ChestX-ray14]

Kumar P, Grewal M, Srivastava M M. Boosted cascaded convnets for multilabel classification of thoracic diseases in chest radiographs[C]//International Conference Image Analysis and Recognition. Springer, Cham, 2018: 546-552.


1. Overview

In this paper

  • propose a cascaded DNN
  • exploit smooth PWE loss

1.1. Multi-Label Classification

1.1.1. Binary Relevance (BR)

  • not account for the interdependence of different class labels
  • BCE loss

1.1.2. Pairwise Error (PWE) Loss

  • maximzie the margin between positive and negative labels within an example

1.2. Architecture



  • weighted sampling

1.3. Details

  • randomly separate 20% for test
  • 6-level cascading both for cross-entropy and PWE loss
  • ReLU, 0.5 dropout between FC
  • SGD 0.1

1.4. Results