1 of 1

Application of Neural Networks for Aiding Diagnosis of Lung Disorders R Chopade1,A Stanam2, S Pawar3

1Indian Institute of Technology, Kharagpur

2Univeristy of Iowa, Iowa City, Iowa

3Department of Computer Science, Claflin University, Orangeburg, South Carolina

Background

Chest X-rays are currently the best available method for diagnosing different lung associated diseases (Parveen et al., 2011). Chest X-ray exam is one of the most frequent and cost-effective medical imaging examinations. However clinical diagnosis of chest X-ray can be challenging, and sometimes believed to be harder than diagnosis via chest CT imaging. Even some promising works have been reported in the past to achieve a clinically relevant computer-aided detection and diagnosis (CAD) in real world medical sites on all data settings of chest X-rays is still very difficult (Speets et al., 2006). With the technology and computing power, the earlier identification of diseases, particularly lung disease can be detected earlier and more accurately, which can save many people as well as reduce the pressure on the system as the health system has not developed in time with the development of the population (Smith-Bindman et al., 2008). We have consistently shown effective usage of supervised and un-supervised machine learning techniques for disease classifications and biomarker predictions (Pawar et al., 2020, 2020, 2021). In this project, we hypothesize the effectiveness of neural networks in prediction (‘Infiltration’, ‘Atelectasis’, ‘Fibrosis’ & ‘Pneumothorax’) and localization ('Cardiomegaly') of specific lung disorders.

Materials and Methods

Counts per class for primary labels after one-hot encoding.

Results

Model architecture used for all the binary classifiers

Loss vs. learning rate plot for “Pneumothorax” binary classifier trained for 10 epochs

Prediction of location of cardiomegaly with bounding boxes

Accuracy plot for “Pneumothorax” binary classifier with constant learning rate, CLR with “triangular” and CLR with “modified triangular2” policies

Plot showing the "Triangular" policy for “Pneumothorax” binary classifier trained for 36 epochs

Image Label

No. of Images before One Hot Encoding

No. of Image before One Hot Encoding

No Finding

60361

60361

Atelectasis

4215

11559

Cardiomegaly

1093

2776

Consolidation

1310

4667

Edema

628

2303

Emphysema

892

2516

Effusion

3955

13317

Fibrosis

727

1686

Infiltration

9547

19894

Mass

2139

5782

Nodule

2705

6331

Pneumothorax

2194

5302

Pneumonia

322

1431

Pleural Thickening

1126

3385

Hernia

110

227

Plot for the "modified triangular2" policy of “Pneumothorax” binary classifier trained for 42 epochs