def process_cam(cam, model_id, model, img_NAC, device):
if model_id == 1: # gating signals of all attention gates of att resnet18_unet were replaced by Grad-CAM
cam1 = np.float32(resize_sitk_2D(cam, (24, 24)))
cam2 = np.float32(resize_sitk_2D(cam, (48, 48)))
cam3 = np.float32(resize_sitk_2D(cam, (96, 96)))
cam4 = np.float32(resize_sitk_2D(cam, (192, 192)))
cam1 = np.expand_dims(np.expand_dims(cam1, axis=0), axis=0)
cam2 = np.expand_dims(np.expand_dims(cam2, axis=0), axis=0)
cam3 = np.expand_dims(np.expand_dims(cam3, axis=0), axis=0)
cam4 = np.expand_dims(np.expand_dims(cam4, axis=0), axis=0)
cam1 = torch.from_numpy(cam1).to(device)
cam2 = torch.from_numpy(cam2).to(device)
cam3 = torch.from_numpy(cam3).to(device)
cam4 = torch.from_numpy(cam4).to(device)
pred = model(img_NAC, cam1, cam2, cam3, cam4)
return pred
elif model_id == 2: # gating signal of 4th attention gate of att resnet18_unet was replaced by Grad-CAM
cam = np.float32(resize_sitk_2D(cam, (192, 192)))
cam = np.expand_dims(np.expand_dims(cam, axis=0), axis=0)
cam = torch.from_numpy(cam).to(device)
pred = model(img_NAC, cam)
return pred
def store_cam(tensor_img_AC, tensor_img_NAC, output_dir, time_stamp,
model_name, slice_id):
grads_val = grad_block[0].cpu().data.numpy().squeeze(0)
fmap = fmap_block[0].cpu().data.numpy().squeeze(0)
cam = gen_cam(fmap, grads_val)
cam = np.float32(resize_sitk_2D(cam, (344, 344)))
tensor_img_NAC = tensor_img_NAC.cpu().data.numpy().squeeze()
tensor_img_NAC = np.float32(resize_sitk_2D(tensor_img_NAC, (344, 344)))
tensor_img_AC = tensor_img_AC.cpu().data.numpy().squeeze()
tensor_img_AC = np.float32(resize_sitk_2D(tensor_img_AC, (344, 344)))
show_cam_on_image(tensor_img_AC, tensor_img_NAC, cam, output_dir,
time_stamp, model_name, slice_id)
print('Attention Map of', str(model_name), 'with slice_id', str(slice_id),
'is saved!')
def load_data(volume_path, slice_id):
volume = sitk.ReadImage(volume_path)
img = volume[:, :, slice_id]
np_img = sitk.GetArrayFromImage(img)
resize_img = resize_sitk_2D(np_img, (384, 384))
transform = transforms.Compose([transforms.ToTensor()])
tensor_img = transform(resize_img)
tensor_img = tensor_img.unsqueeze(0)
return tensor_img
def process_cam(cam, model_id, model, img_NAC, device):
if model_id == 1:
cam1 = np.float32(resize_sitk_2D(cam, (48, 48)))
cam2 = np.float32(resize_sitk_2D(cam, (96, 96)))
cam3 = np.float32(resize_sitk_2D(cam, (192, 192)))
cam4 = np.float32(resize_sitk_2D(cam, (384, 384)))
cam1 = np.expand_dims(np.expand_dims(cam1, axis=0), axis=0)
cam2 = np.expand_dims(np.expand_dims(cam2, axis=0), axis=0)
cam3 = np.expand_dims(np.expand_dims(cam3, axis=0), axis=0)
cam4 = np.expand_dims(np.expand_dims(cam4, axis=0), axis=0)
cam1 = torch.from_numpy(cam1).to(device)
cam2 = torch.from_numpy(cam2).to(device)
cam3 = torch.from_numpy(cam3).to(device)
cam4 = torch.from_numpy(cam4).to(device)
pred = model(img_NAC, cam1, cam2, cam3, cam4)
return pred
elif model_id == 2:
cam = np.float32(resize_sitk_2D(cam, (384, 384)))
cam = np.expand_dims(np.expand_dims(cam, axis=0), axis=0)
cam = torch.from_numpy(cam).to(device)
pred = model(img_NAC, cam)
return pred
elif model_id == 3:
cam = np.float32(resize_sitk_2D(cam, (48, 48)))
cam = np.expand_dims(np.expand_dims(cam, axis=0), axis=0)
cam = torch.from_numpy(cam).to(device)
pred = model(img_NAC, cam)
return pred
elif model_id == 4:
cam1 = np.float32(resize_sitk_2D(cam, (48, 48)))
cam2 = np.float32(resize_sitk_2D(cam, (96, 96)))
cam1 = np.expand_dims(np.expand_dims(cam1, axis=0), axis=0)
cam2 = np.expand_dims(np.expand_dims(cam2, axis=0), axis=0)
cam1 = torch.from_numpy(cam1).to(device)
cam2 = torch.from_numpy(cam2).to(device)
pred = model(img_NAC, cam1, cam2)
return pred
elif model_id == 5:
cam1 = np.float32(resize_sitk_2D(cam, (48, 48)))
cam2 = np.float32(resize_sitk_2D(cam, (96, 96)))
cam3 = np.float32(resize_sitk_2D(cam, (192, 192)))
cam1 = np.expand_dims(np.expand_dims(cam1, axis=0), axis=0)
cam2 = np.expand_dims(np.expand_dims(cam2, axis=0), axis=0)
cam3 = np.expand_dims(np.expand_dims(cam3, axis=0), axis=0)
cam1 = torch.from_numpy(cam1).to(device)
cam2 = torch.from_numpy(cam2).to(device)
cam3 = torch.from_numpy(cam3).to(device)
pred = model(img_NAC, cam1, cam2, cam3)
return pred