Noisy Image Classification

Overall: The paper proposes a new approach in which removal of noise is no longer necessary. However, they should have compared their results with a removed noise test set

Medical image datasets

Increasing in size and availability

Training for deep learning models yielded better accuracy scores

Labeling can now be automated

Use of Natural Language Processing (NLP)

Prone to noises

Use of Chest X-ray 14 dataset

Known for noisy sample

Contribution: estimating performance of clean test set from noisy test set

Deep Neural Networks (DNN)

Depend on dataset being fed into models

Challenges

Difficult/expensive to extract

Time-consuming to label

Even specialized radiologists find it difficult

Challenges

Noisy samples lead to overfitting

Noisy samples lead to low accuracy performance

Methodology

Experiment and Results

Classification of multiple class images

Use of early learning regularization (ELR)

Noisy label learning

Robust loss function

Transition matrices

Sample selection

Cannot be used for multiple-class labels

Labels can be extracted through NLP

Can be unreliable

Prone to errors

Estimating accuracy of clean test set

Lower bound estimation

Function depending on several variables

Algorithm to regularize cross-entropy loss

For boosting clean gradients

For dampening noisy gradients

Accuracy of noisy test set

Value of delta mapped from 0 to 1

Size of test set

Noise rate

Noise transition matrix

Dataset consists of 15 classes

Initial 14 classes

Addition of "No Label" classs

ELR outperformed several state-of-the-art models

Lower bounds for clean test set accuracy

Lower than noisy test set accuracy

Note: Better if they just removed the noise (?)

Overall: The paper proposes a new approach in which removal of noise is no longer necessary. However, they should have compared their results with a removed noise test set

Critique: High accuracy is necessary especially for highly sensitive information such as medical images. If deployed in a real setting, their approach, although novel, should yield higher results.

Liu, F., Tian, Y., Cordeiro, F. R., Belagiannis, V., Reid, I., and Carneiro, G. Noisy label learning for large-scale medical image classification, 2021.