Quick Start

This tutorial introduces you to the basics of deep_texture_histology workflow.

DTR calculation

import deeptexture as dt
import glob

# create DTR object
dtr_obj = dt.DTR(arch='vgg', layer='block4_conv3', dim=1024)

# calculate DTR for one image file
imgfile = './example/0KNpXTsaNix2T6.jpg'
dtr = dtr_obj.get_dtr(imgfile)

# calculate mean DTR for unrotated and rotated image file
imgfile = './example/0KNpXTsaNix2T6.jpg'
dtr = dtr_obj.get_dtr(imgfile)
dtr_rot = dtr_obj.get_dtr(imgfile, angle=[0, 90])

# calculate DTR for one image object
imgfile = './example/0KNpXTsaNix2T6.jpg'
dtr = dtr_obj.get_dtr(imgfile)
img = Image.open(imgfile)
dtr = dtr_obj.get_dtr(img)

# calculate DTR for one image object
imgfile = './example/0KNpXTsaNix2T6.jpg'
dtr = dtr_obj.get_dtr(imgfile)
img_np = np.array(Image.open(imgfile))
dtr = dtr_obj.get_dtr(img_np)

# calculate DTRs for multiple images
imgfiles = glob.glob("./example/*.jpg")
dtrs = dtr_obj.get_dtr_multifiles(imgfiles)

# calculate DTRs for multiple unrotated and rotated images
imgfiles = glob.glob("./example/*.jpg")
dtrs = dtr_obj.get_dtr_multifiles(imgfiles, angle=[0, 90])

Plot DTR distribution

1. Plot images

   from deeptexture import plt_dtr
   
   imgfiles = ['./example/0KNpXTsaNix2T6.jpg',
               './example/0zpw5BpfoiRX9m.jpg',
               './example/11303cFBFubBwG.jpg',
               './example/19HDnAR7OBcA1d.jpg',]
   text = ['01','02','03','04',]
   
   dtrs = dtr_obj.get_dtr_multifiles(imgfiles, angle=[0, 90])
   
   #Principal component analysis is applied to DTRs and PC1 and PC2 is shown.
   X_emb = plt_dtr.plt_dtr_image(dtrs, imgfiles, method="pca", x1=1, x2=2, text=text)
   
   #Calculated embedding can be the input.
   _ = plt_dtr.plt_dtr_image(X_emb, imgfiles, text=text)
   

2. Plot attributes

   from deeptexture import plt_dtr
   
   imgfiles = ['./example/0KNpXTsaNix2T6.jpg',
               './example/0zpw5BpfoiRX9m.jpg',
               './example/11303cFBFubBwG.jpg',
               './example/19HDnAR7OBcA1d.jpg',]
   attr =  ['cancer',
           'cancer',
           'normal',
           'normal',]
   text = ['01','02','03','04',]
   
   dtrs = dtr_obj.get_dtr_multifiles(imgfiles, angle=[0, 90])
   
   #Principal component analysis is applied to DTRs and PC1 and PC2 is shown.
   df = plt_dtr.plt_dtr_attr(dtrs, attr, method="pca", x1=1, x2=2, text=text)
   
   #Calculated embedding can be the input.
   _ = plt_dtr.plt_dtr_attr(X_emb, attr, text=text)
   

Content-based image retrieval

DTR enables quick and accurate retrieval of histologically similar images using CBIR module. You can create original database and save the files in the specified directory.

1. Create CBIR database.

   import pandas as pd
   
   import deeptexture as dt
   from deeptexture import cbir
   
   # create DTR object
   dtr_obj = dt.DTR(arch='vgg', layer='block4_conv3', dim=1024)
   
   # create CBIR object
   cbir_obj = cbir.CBIR(dtr_obj, project='DB', working_dir='CBIR')
   
   # create CBIR database
   imgfiles = ['./example/0KNpXTsaNix2T6.jpg',
               './example/0zpw5BpfoiRX9m.jpg',
               './example/11303cFBFubBwG.jpg',
               './example/19HDnAR7OBcA1d.jpg',]
   patients = ['01',
               '02',
               '03',
               '04',]
   attr =  ['cancer',
           'cancer',
           'normal',
           'normal',]
   df_attr = pd.DataFrame({'imgfile': imgfiles,
                           'patient': patients,
                           'tissue': attr})
   
   cbir_obj.create_db(df_attr, img_attr='imgfile', save=True)
   

2. Search similar histology images.

   # search the most similar images (top two)
   qimgfile = "./example/3t0U7aBXRY9v1V.jpg"
   cbir_obj.search(qimgfile, img_attr='imgfile', case_attr='patient', n=2)
   

Supervised learning model

import deeptexture as dt
from deeptexture import ml

imgfiles = ['./example/0KNpXTsaNix2T6.jpg',
            './example/0zpw5BpfoiRX9m.jpg',
            './example/11303cFBFubBwG.jpg',
            './example/19HDnAR7OBcA1d.jpg',]
patients = ['01',
            '02',
            '03',
            '04',]
attr =  ['cancer',
         'cancer',
         'normal',
         'normal',]
df_attr = pd.DataFrame({'imgfile': imgfiles,
                        'patient': patients,
                        'tissue': attr})

dtrs = dtr_obj.get_dtr_multifiles(imgfiles, angle=[0, 90])

# create ml object
ml_obj = ml.ML(dtrs, df_attr.imgfile)

# create a model to classify image into tissue type
y = df_attr.tissue
cases = df_attr.patient

lr = ml_obj.fit_eval(y, cases)