Pneumonia Detection using machine learning - with PyTorch

Last update: Jul 07, 2022

Overview

Pneumonia Detection

Pneumonia Detection using machine learning.

Training was done in colab:

DEMO:

Result (Confusion Matrix):

Data

I uploaded my dataset to kaggle I used a modified version of this dataset from kaggle. Instead of NORMAL and PNEUMONIA I split the PNEUMONIA dataset to BACTERIAL PNUEMONIA and VIRAL PNEUMONIA. This way the data is more evenly distributed and I can distinguish between viral and bacterial pneumonia. I also combined the validation dataset with the test dataset because the validation dataset only had 8 images per class.

This is the resulting distribution:

Processing and Augmentation

I resized the images to 150x150 and because some images already were grayscale I also transformed all the images to grayscale.

Additionaly I applied the following transformations/augmentations on the training data:

transforms.Resize((150, 150)),
transforms.Grayscale(),
transforms.ToTensor(),
transforms.RandomHorizontalFlip(),
transforms.RandomVerticalFlip(),
transforms.RandomRotation(45)

and those transformations on the test data:

transforms.Resize((150, 150)),
transforms.Grayscale(),
transforms.ToTensor(),

This is the resulting data:

I also used one-hot encoding for the labels!

Model

----------------------------------------------------------------
        Layer (type)               Output Shape         Param #
================================================================
            Conv2d-1         [-1, 16, 148, 148]             160
              ReLU-2         [-1, 16, 148, 148]               0
       BatchNorm2d-3         [-1, 16, 148, 148]              32
            Conv2d-4         [-1, 16, 146, 146]           2,320
              ReLU-5         [-1, 16, 146, 146]               0
       BatchNorm2d-6         [-1, 16, 146, 146]              32
         MaxPool2d-7           [-1, 16, 73, 73]               0
            Conv2d-8           [-1, 32, 71, 71]           4,640
              ReLU-9           [-1, 32, 71, 71]               0
      BatchNorm2d-10           [-1, 32, 71, 71]              64
           Conv2d-11           [-1, 32, 69, 69]           9,248
             ReLU-12           [-1, 32, 69, 69]               0
      BatchNorm2d-13           [-1, 32, 69, 69]              64
        MaxPool2d-14           [-1, 32, 34, 34]               0
           Conv2d-15           [-1, 64, 32, 32]          18,496
             ReLU-16           [-1, 64, 32, 32]               0
      BatchNorm2d-17           [-1, 64, 32, 32]             128
           Conv2d-18           [-1, 64, 30, 30]          36,928
             ReLU-19           [-1, 64, 30, 30]               0
      BatchNorm2d-20           [-1, 64, 30, 30]             128
        MaxPool2d-21           [-1, 64, 15, 15]               0
           Conv2d-22          [-1, 128, 13, 13]          73,856
             ReLU-23          [-1, 128, 13, 13]               0
      BatchNorm2d-24          [-1, 128, 13, 13]             256
           Conv2d-25          [-1, 128, 11, 11]         147,584
             ReLU-26          [-1, 128, 11, 11]               0
      BatchNorm2d-27          [-1, 128, 11, 11]             256
        MaxPool2d-28            [-1, 128, 5, 5]               0
          Flatten-29                 [-1, 3200]               0
           Linear-30                 [-1, 4096]      13,111,296
             ReLU-31                 [-1, 4096]               0
          Dropout-32                 [-1, 4096]               0
           Linear-33                 [-1, 4096]      16,781,312
             ReLU-34                 [-1, 4096]               0
          Dropout-35                 [-1, 4096]               0
           Linear-36                    [-1, 3]          12,291
          Softmax-37                    [-1, 3]               0
================================================================
Total params: 30,199,091
Trainable params: 30,199,091
Non-trainable params: 0
----------------------------------------------------------------
Input size (MB): 0.09
Forward/backward pass size (MB): 27.95
Params size (MB): 115.20
Estimated Total Size (MB): 143.24
----------------------------------------------------------------

Visualization using Streamlit

The webapp is not hosted because the model is too large. I'd have to host it on a server. This is just to visualize.

Pneumonia Detection using machine learning - with PyTorch

Related tags

Overview

Pneumonia Detection

Data

Processing and Augmentation

Model

Visualization using Streamlit

Owner

Wilhelm Berghammer

Segment axon and myelin from microscopy data using deep learning

Pixel-level Crack Detection From Images Of Levee Systems : A Comparative Study

Code for "PV-RAFT: Point-Voxel Correlation Fields for Scene Flow Estimation of Point Clouds", CVPR 2021

HybVIO visual-inertial odometry and SLAM system

Code for A Volumetric Transformer for Accurate 3D Tumor Segmentation

Code for our paper "MG-GAN: A Multi-Generator Model Preventing Out-of-Distribution Samples in Pedestrian Trajectory Prediction" published at ICCV 2021.

Code for our NeurIPS 2021 paper 'Exploiting the Intrinsic Neighborhood Structure for Source-free Domain Adaptation'

This initial strategy was developed specifically for larger pools and is based on taking a moving average and deriving Bollinger Bands to create a projected active liquidity range.

Chainer Implementation of Semantic Segmentation using Adversarial Networks

Fantasy Points Prediction and Dream Team Formation

Real-time multi-object tracker using YOLO v5 and deep sort

ElegantRL is featured with lightweight, efficient and stable, for researchers and practitioners.

Volumetric parameterization of the placenta to a flattened template

UMich 500-Level Mobile Robotics Course

An official implementation of "Exploiting a Joint Embedding Space for Generalized Zero-Shot Semantic Segmentation" (ICCV 2021) in PyTorch.

Cognition-aware Cognate Detection

Sarus implementation of classical ML models. The models are implemented using the Keras API of tensorflow 2. Vizualization are implemented and can be seen in tensorboard.

Code for "ATISS: Autoregressive Transformers for Indoor Scene Synthesis", NeurIPS 2021

Tensorflow solution of NER task Using BiLSTM-CRF model with Google BERT Fine-tuning And private Server services

.NET bindings for the Pytorch engine