Histopathologic Cancer Detector

Python Jupyter Notebook leveraging Transfer Learning and Convolutional Neural Networks implemented with Keras.

Part of the Kaggle competition.

Submitted Kernel with 0.958 LB score.

Check out corresponding Medium article:

Histopathologic Cancer Detector - Machine Learning in Medicine

Data

Dataset: Link

Description: Binary classification whether a given histopathologic image contains a tumor or not.

Training: 153k (0.9) images

Validation: 17k (0.1) images

Testing: 57.5k images

Model

__________________________________________________________________________________________________
Layer (type)                    Output Shape         Param #     Connected to                     
==================================================================================================
input_1 (InputLayer)            (None, 96, 96, 3)    0                                            
__________________________________________________________________________________________________
xception (Model)                (None, 3, 3, 2048)   20861480    input_1[0][0]                    
__________________________________________________________________________________________________
NASNet (Model)                  (None, 3, 3, 1056)   4269716     input_1[0][0]                    
__________________________________________________________________________________________________
global_average_pooling2d_1 (Glo (None, 2048)         0           xception[1][0]                   
__________________________________________________________________________________________________
global_average_pooling2d_2 (Glo (None, 1056)         0           NASNet[1][0]                     
__________________________________________________________________________________________________
concatenate_5 (Concatenate)     (None, 3104)         0           global_average_pooling2d_1[0][0] 
                                                                 global_average_pooling2d_2[0][0] 
__________________________________________________________________________________________________
dropout_1 (Dropout)             (None, 3104)         0           concatenate_5[0][0]              
__________________________________________________________________________________________________
dense_1 (Dense)                 (None, 1)            3105        dropout_1[0][0]                  
==================================================================================================
Total params: 25,134,301
Trainable params: 25,043,035
Non-trainable params: 91,266
__________________________________________________________________________________________________

Training

Results

Kaggle score: 0.958

Author

Greg (Grzegorz) Surma

PORTFOLIO

GITHUB

BLOG