shanmukh05 / Pneumonia-Detection-Challenge

classify CXRs with pneumonia from their normal CXR counterparts, using machine learning and computer vision techniques.

Pneumonia-Detection-Challenge - DPhi

Check-out the competition here

Final Leaderboard - 5th position in Private LB and 7th in Public LB

Problem Statement

Classify chest X-rays(CXRs) with pneumonia from their normal CXR counterparts, using machine learning and computer vision techniques.

Dataset

Dataset and pretrained models are uploaded in Kaggle

Classes

• Normal
• Pneumonia

Training Data

• Normal : 1280
• Pneumonia : 1145
• Distribution :

Test Data

• 606 images

Training

• As the given dataset contains very less instances, training from scratch will not given best results. So I used following pretrained models

  • ChexNet - keras Implementation
  • ChexNet - VGG

• TPU as Accelerator

• GPU as Accelerator

• Best Result

• Submission Notebook

Best Results

• Metric : Accuracy
• Best Results are stored in experiments.xlsx. (Versions denotes the version in Kaggle notebook.)
• ChexNet Implementation with freezing bottom layers gave best results (DenseNet121 Implementation)
• Following is the model with highest score (Private LB : 83.16, Public LB : 80.61).

    model = tf.keras.applications.DenseNet121(weights= "imagenet",
                                    include_top=False,
                                    input_shape=(HEIGHT,WIDTH,CHANNELS), pooling="avg")
    predictions = tf.keras.layers.Dense(14, activation='sigmoid', name='predictions')(model.output)
    model = tf.keras.Model(inputs=model.input, outputs=predictions)
    
    model.load_weights("../input/pneumonia-classification-challenge/pretrained.h5")
    model = tf.keras.Model(model.input, model.layers[-2].output)
    
    x = tf.keras.layers.Dense(512, activation = "relu")(model.output)
    x = tf.keras.layers.Dropout(0.3)(x)
    x = tf.keras.layers.Dense(128, activation = "relu")(x)
    x = tf.keras.layers.Dense(64, activation = "relu")(x)
    outputs = tf.keras.layers.Dense(2, activation = "softmax", dtype = tf.float32)(x)
    model = tf.keras.Model(model.input,outputs)
    
    for layer in model.layers[:-14]:
        layer.trainble = False