Exploiting transfer learning methods to try and classify X-ray chest Images into normal(healthy) vs abnormal(sick)
we will see the performance of transfer learning using the official pre-trained model offered by Google (INCEPTION-RESNET-V2 MODEL), which can be found in TensorFlow’s model library
In this little/first try we will be retraining the last layer of inception v2 of google to classify the images using adam optimizer and learning rate decay
python 3 tensorflow = 1.0.1 matplotlib lxml
model used : INCEPTION-RESNET-V2
learning rate : 0.0001 with a decay factor of 0.7 each 2 epochs
batch size : 16
number of epochs : 30
Streaming Accuracy : 68.70 %
Recall : coming soon
Precision : coming soon
In the data folder (cd data/) :
1 - Use python get_data.py to download scrapped image data from openi.nlm.nih.gov. It has a large base of Xray,MRI, CT scan images publically available.Specifically Chest Xray Images have been scraped.The images will be downloaded and saved in images/ and the labels in data_new.json (it might take a while)
Some info about the dataset :
Total number of Images : 7469
The classes with most occurence in the dataset:
("normal", 2696)
("No Indexing", 172)
("Lung/hypoinflation", 88)
("Thoracic Vertebrae/degenerative/mild", 55)
("Thoracic Vertebrae/degenerative", 44)
("Spine/degenerative/mild", 36)
("Spine/degenerative", 35)
("Spondylosis/thoracic vertebrae", 33)
("Granulomatous Disease", 32)
("Cardiomegaly/mild", 32)
2 - Use python gen_data.py to sort labels into Normal/Abnormal classes, generate full path to coresponding Images and write them to data.txt
number of normal chest Images(healthy people) 2696:
number of abnormal chest Images(sick people) 4773:
3 - Use python convert_to_tf_records.py to generate tf records of the data.
Download the Pre-trained inception model in here and unzip it in ckpt/ folder.
Use python train.py to start the training !(trained model will be saved in logs/)
Use python evaluate.py to run evaluation using the model saved in logs/(metric : streaming accuracy over all mini batches)