Pneumothorax demo

from fastai.vision.all import *
from fastai.medical.imaging import *
import torchvision
import torchvision.transforms as t

pneumothorax_source = untar_data(URLs.SIIM_SMALL)

items = get_dicom_files(pneumothorax_source/f"train/")

df = pd.read_csv(pneumothorax_source/f"labels.csv")
df.head()

	file	label
0	train/No Pneumothorax/000000.dcm	No Pneumothorax
1	train/Pneumothorax/000001.dcm	Pneumothorax
2	train/No Pneumothorax/000002.dcm	No Pneumothorax
3	train/Pneumothorax/000003.dcm	Pneumothorax
4	train/Pneumothorax/000004.dcm	Pneumothorax

df_pneumo = df[df['label']=='Pneumothorax']

pneumothorax = DataBlock(blocks=(ImageBlock(cls=PILDicom), MaskBlock(codes=['bg','pneumo'])),
                   get_x=lambda x:pneumothorax_source/f"{x[0]}",
                   get_y=lambda x:'example/siim/'+Path(x[0]).stem+'.png',
                   batch_tfms=aug_transforms(do_flip=False,p_affine=0,p_lighting=0,size=224))

dls = pneumothorax.dataloaders(df_pneumo.values, num_workers=0, bs=16)

dls.show_batch(max_n=16, vmin=0)

# just for testing - you can skip this when running interactively
learn = unet_learner(dls, resnet34)

learn.fine_tune(1)

epoch	train_loss	valid_loss	time
0	1.382017	0.309487	01:04

epoch	train_loss	valid_loss	time
0	0.167802	0.111701	00:51

# this is skipped when testing, instead a single epoch is done (not sufficient for overfitting)
learn = unet_learner(dls, resnet34)
learn.fine_tune(8)

epoch	train_loss	valid_loss	time
0	2.200016	0.416453	01:05

epoch	train_loss	valid_loss	time
0	0.195540	0.105748	00:51
1	0.148728	0.141623	00:51
2	0.126741	0.176495	00:54
3	0.116757	0.106115	00:57
4	0.106740	0.093054	02:16
5	0.098731	0.140762	01:49
6	0.092059	0.120493	02:27
7	0.086470	0.116142	01:48

learn.show_results(ds_idx=0,max_n=9,vmin=0)

preds = learn.get_preds(ds_idx=0)

plt.imshow(preds[1][2])

<matplotlib.image.AxesImage>

preds[0].argmax(dim=1).sum()
#plt.imshow(preds[0].argmax(dim=1)[0])

TensorBase(0)

learn.fine_tune(8)

epoch	train_loss	valid_loss	time
0	0.061920	2.473048	01:03

epoch	train_loss	valid_loss	time
0	0.314660	0.080977	01:54
1	0.208264	0.088037	00:55
2	0.160613	0.078349	00:54
3	0.134872	0.086172	00:52
4	0.113790	0.082046	00:52
5	0.100933	0.077185	00:53
6	0.090635	0.089535	01:10
7	0.083492	0.089540	01:10

learn.recorder.plot_loss()

preds = learn.get_preds(ds_idx=0)
preds[0].argmax(dim=1).sum()

TensorBase(0)

fig, axs = plt.subplots(5,2,figsize=(8,16))
for i in range(5):
    axs[i,0].imshow(preds[1][i])
    axs[i,1].imshow(preds[0][i][1])
    print(preds[0][i][1].max())

TensorBase(0.3102)
TensorBase(0.3205)
TensorBase(0.3076)
TensorBase(0.2694)
TensorBase(0.0728)

learn.fit_one_cycle(8)

epoch	train_loss	valid_loss	time
0	0.043738	0.085497	01:06
1	0.041399	0.126955	00:53
2	0.039557	0.166106	00:52
3	0.038748	0.160928	00:55
4	0.036368	0.123756	00:52
5	0.034367	0.111612	00:52
6	0.032606	0.130679	00:52
7	0.030831	0.122198	00:53

preds = learn.get_preds(ds_idx=0)
preds[0].argmax(dim=1).sum()

TensorBase(14126)

learn.fit_one_cycle(20)

epoch	train_loss	valid_loss	time
0	0.023082	0.114249	00:56
1	0.022736	0.110707	00:53
2	0.021745	0.209588	00:52
3	0.022693	0.105819	01:18
4	0.024324	0.151829	00:49
5	0.024436	0.093544	00:49
6	0.024203	0.123490	00:48
7	0.023524	0.166477	00:50
8	0.022349	0.262088	00:55
9	0.021310	0.254958	01:41
10	0.020904	0.199529	02:09
11	0.020301	0.199600	00:58
12	0.019917	0.186603	01:07
13	0.019124	0.135860	00:52
14	0.018472	0.152205	00:49
15	0.017754	0.178598	00:52
16	0.017116	0.187399	00:56
17	0.016518	0.185505	00:56
18	0.015890	0.184788	00:57
19	0.015337	0.181746	00:57

preds = learn.get_preds(ds_idx=0)
preds[0].argmax(dim=1).sum()

TensorBase(41177)

fig, axs = plt.subplots(5,3,figsize=(12,16))
for i in range(5):
    axs[i,0].imshow(preds[1][i], interpolation="nearest")
    axs[i,1].imshow(preds[0][i][1])
    axs[i,2].imshow(preds[0].argmax(dim=1)[i], interpolation="nearest")

learn.save('pneumothorx')

Path('models/pneumothorx.pth')

We managed to overfit

learn.load('pneumothorx')

<fastai.learner.Learner>

from misas.core import *
from fastai.vision.all import * #open_mask, Image, ImageSegment
import pydicom

def read_dcm(file):
    ds = pydicom.dcmread(file)
    img = Tensor(ds.pixel_array.astype(np.int16))
    img = img/img.max()
    transform = t.ToPILImage()
    img = transform (torch.stack([img, img, img]))
    return img

learn.predict(learn.dls.train_ds.items[1])[0].show(vmin=0)

<AxesSubplot:>

fname = df_pneumo.iloc[3]['file']
img = lambda: read_dcm(pneumothorax_source/fname)
trueMask = lambda: Image.open('example/siim/'+Path(fname).stem+'.png') #open_mask('example/siim/'+Path(fname).stem+'.png')

dcm = img()

dcm
#to_image(dcm.data)

pred = learn.predict(PILDicom(dcm))[1]
plt.imshow(pred)

<matplotlib.image.AxesImage>

class Fastai2_wrapper:
    def __init__(self, model):
        self.model = model
        self.model.cbs = L([])
    
    def imageToPILDicom(self, image):
        return PILDicom(image)
    
    def prepareSize(self, imageOrMask):
        return imageOrMask.resize((224,224))
    
    def predict(self, image):
        image = self.imageToPILDicom(image)
        pred = self.model.predict(image)[0]#(in_image)
        return Image.fromarray(np.array(pred).astype(np.uint8))

model = Fastai2_wrapper(learn)

plt.imshow(model.predict(img()))

<matplotlib.image.AxesImage>

plot_series(get_rotation_series(img(), model))

result = eval_rotation_series(img(), trueMask(), model, start=-180, end=180, components=['bg','pneumo'])

plt.plot(result['deg'],result['pneumo'])