Skip to content

Latest commit

 

History

History
238 lines (212 loc) · 9.18 KB

README.md

File metadata and controls

238 lines (212 loc) · 9.18 KB
Raw

Multiclass Disease Classification using Modified CNN and Segmented Chest X-ray

Multiclass classification is always a challenge for many researcher. It is seen that for image classification task mostly for binary classifcation many researcher have achieved accuracy of more than 99%. Starting with the cat and dog classification,image classification technique is very much utilised in the field of medical images, for various diagnosis and prediction.

Background

We will be modifing and testing various model which have achieved higher accuracy for multiclass classification problem to achieve good result for both segmented data as well as normal image for 4 class classification consisting of Pneumonia,Tuberculosis,Normal,Covid-19.

Method

We have collected Chest X-Ray(CXR) images of various dimensions with label Pneumonia,Normal,Covid-19,Tuberculosis, and resize it to 128x128 pixel grayscale images. We have developed two types of dataset:

  • Segmented Image Dataset
  • Normal Image Dataset

Segmented Image Dataset

We have used U-Net model for the prediction of mask for selecting only lung portion. Although Accuracy of U-Net model was itself very low so we get very few clean segmented mask. This may be the reason we are getting low accuracy for the segmented data.

Why we want Mask?

Mask will assure us that our model is learning on the right data of the CXR, and build the confidence among user, It can also be helpful for skipping model computation at certain pixel of the images.

Normal Image Dataset

It is the simple CXR images which is flatten to form ndarray.

target are kept categorical

Total Images
Training 6326
Validating 38
Testing 771

Prepared Dataset Link: CXR DATA for Multiclass Classfification

Analysis

Model Architecture

VGG-13

Model Performance

Without Segmentation

Metric Table

Model Training Testing
Accuracy Precision Recall F1-score Accuracy Precision Recall F1-score
VGG-13 0.9512 0.9531 0.9485 0.9508 0.8560 0.8620 0.8508 0.8564
AlexNet 0.9683 0.9709 0.9658 0.9683 0.8197 0.8337 0.8132 0.8233
MobileNet 0.9918 0.9924 0.9915 0.9919 0.8457 0.8449 0.8405 0.8427
Modified-DarkCovidNet 0.9813 0.9814 0.9810 0.9812 0.8313 0.8344 0.8300 0.8322

Confusion Metrics

VGG-13 AlexNet MobileNet
Modified-DarkCovidNet AlexNet MobileNet

With segmentation

Metric Table

Model Training Testing
Accuracy Precision Recall F1-score Accuracy Precision Recall F1-score
VGG-13 0.9317 0.9367 0.9287 0.9327 0.7613 0.7668 0.7549 0.7608
AlexNet 0.9625 0.9652 0.9594 0.9623 0.7691 0.7781 0.7639 0.7709
MobileNet 0.9344 0.9389 0.9301 0.9345 0.7756 0.7795 0.7704 0.7749
Modified-DarkCovidNet 0.9483 0.9497 0.9473 0.9485 0.7924 0.7935 0.7924 0.7929

Confusion Metrics

VGG-13 AlexNet MobileNet
Modified-DarkCovidNet AlexNet MobileNet

Comparison