def update_choice(self, **kwargs):
self.liver = kwargs.get("liver", self.liver)
self.label = kwargs.get("label", self.label)
self.mask = array_kits.merge_labels(self.mask_,
[0, [1, 2]] if self.liver else [0, 2]).astype(np.int8) * 2
self.pred = array_kits.merge_labels(self.pred_,
[0, [1, 2]] if self.liver else [0, self.label]).astype(np.int8) * 2
def dump_fp_bbox_from_prediction(label_dirs, pred_dir):
pred_dir = Path(pred_dir)
save_path = pred_dir.parent / "bboxes-{}.pkl".format(pred_dir.parent.name)
all_bboxes = {}
counter = 0
for pred_path in sorted(pred_dir.glob("prediction-*.nii.gz")):
print(pred_path.name)
lab_file = pred_path.stem.replace("prediction", "segmentation")
lab_path = misc.find_file(label_dirs, lab_file)
result = array_kits.merge_labels(
nii_kits.nii_reader(pred_path)[1], [0, 2])
reference = array_kits.merge_labels(
nii_kits.nii_reader(lab_path)[1], [0, 2])
fps, tps = array_kits.find_tp_and_fp(result, reference)
for x in fps:
print(x, [x[3] - x[0], x[4] - x[1], x[5] - x[2]])
print()
counter += len(fps)
for x in tps:
print(x, [x[3] - x[0], x[4] - x[1], x[5] - x[2]])
print("#" * 80)
all_bboxes[int(pred_path.stem.replace(".nii", "").split("-")[-1])] = {
"fps": fps,
"tps": tps
}
print("FPs: {}".format(counter))
with save_path.open("wb") as f:
pickle.dump(all_bboxes, f, pickle.HIGHEST_PROTOCOL)
def dump_all_tumor_bbox():
save_file = Path(__file__).parent.parent / "data/LiTS/tumor_summary.csv"
disc = ndi.generate_binary_structure(3, connectivity=1)
info = collections.defaultdict(list)
for lits in LiTS_ROOTS:
for mask_path in sorted(lits.glob("segmentation-*.nii")):
hdr, mask = nii_kits.nii_reader(mask_path)
sx, sy, sz = hdr["srow_x"][0], hdr["srow_y"][1], hdr["srow_z"][2]
voxel = abs(sx * sy * sz)
mask = array_kits.merge_labels(mask, [0, 2])
labeled, num_obj = ndi.label(mask, disc)
objs = ndi.find_objects(labeled)
for i, obj in enumerate(objs):
bbox = array_kits.slices_to_bbox(obj)
area = np.sum(labeled == i + 1) * voxel
info["PID"].append(mask_path.name)
info["TID"].append(i)
info["min_x"].append(bbox[2])
info["min_y"].append(bbox[1])
info["min_z"].append(bbox[0])
info["max_x"].append(bbox[5])
info["max_y"].append(bbox[4])
info["max_z"].append(bbox[3])
info["area/cc"].append(area)
info["num_slices"].append(bbox[3] - bbox[0])
print(mask_path.name, i, bbox, area)
pd.DataFrame(data=info).to_csv(str(save_file))
def update_case(self, case_path, **kwargs):
self.gt = self.mask = self.mask_ = None
self.pred = self.pred_ = None
case_path = "prediction-{}.nii.gz".format(int(case_path.split("-")[1]))
if Path(case_path).name.endswith(".nii.gz"):
ori_file = find_file(self.data_dirs, case_path.replace("prediction", "volume")
.replace(".nii.gz", ".nii"))
lab_file = find_file(self.data_dirs, case_path.replace("prediction", "segmentation")
.replace(".nii.gz", ".nii"))
pred_file = find_file([self.pred_dir], case_path)
reader = nii_kits.nii_reader
else:
raise ValueError("Wrong prediction name: {}".format(case_path))
if self.liver_range is None and self.bbox_file.exists():
with self.bbox_file.open("rb") as f:
self.liver_range = pickle.load(f)
self.meta, self.gt = reader(ori_file)
self.pred_ = reader(pred_file)[1].astype(np.int8)
self.mask_ = reader(lab_file)[1].astype(np.int8)
self.shape = self.gt.shape
if self.liver_range is not None:
self.bb = self.liver_range[ori_file.name.split(".")[0]][0]
ranges = slice(self.get_min_idx(), self.get_max_idx() + 1)
self.gt = self.gt[ranges]
self.pred_ = self.pred_[ranges]
self.mask_ = self.mask_[ranges]
np.clip(self.gt, -100, 400, self.gt)
self.gt = ((self.gt + 100) * (255 / 500)).astype(np.uint8)
self.liver = kwargs.get("liver", True)
self.label = kwargs.get("label", 2)
if self.liver:
self.mask = array_kits.merge_labels(self.mask_, [0, [1, 2]]).astype(np.int8) * 2
self.pred = array_kits.merge_labels(self.pred_, [0, [1, 2]]).astype(np.int8) * 2
else:
self.mask = array_kits.merge_labels(self.mask_, [0, 2]).astype(np.int8) * 2
self.pred = array_kits.merge_labels(self.pred_, [0, self.label]).astype(np.int8) * 2
assert self.gt.shape == self.pred.shape and self.gt.shape == self.mask.shape, \
"gt: {}, mask: {}, pred: {}".format(self.gt.shape, self.mask.shape, self.pred.shape)
def run(mgr):
_, temp = nii_kits.nii_reader(
Path(__file__).parent / "model_dir/016_osmn_in_noise"
"/prediction/prediction-113.nii.gz")
temp = arr_ops.merge_labels(temp, [0, 2])
temp = temp[arr_ops.bbox_to_slices(mgr.bbox)]
temp = ndi.zoom(
temp,
[1, mgr.shape[0] / temp.shape[1], mgr.shape[1] / temp.shape[2]],
order=0)[n:]
for x in np.concatenate((temp, np.flip(temp, axis=0)), axis=0):
yield x
def convert_to_tp_dataset(dataset,
k_split=5,
folds_file="k_folds.txt",
seed=None,
align=1,
padding=0,
min_bbox_shape=None,
prefix="cls-0tp"):
file_names = get_lits_list(dataset, False)
num_images = len(file_names)
k_folds = read_or_create_k_folds(
Path(__file__).parent.parent / "data/LiTS/{}".format(folds_file),
file_names, k_split, seed)
LiTS_records = _get_lits_records_dir()
label_reader = SubVolumeReader(np.uint8, extend_channel=False)
counter = 1
for i, fold in enumerate(k_folds):
output_filename = LiTS_records / "{}-{}-of-{}.tfrecord".format(
prefix, i + 1, k_split)
with tf.io.TFRecordWriter(str(output_filename)) as writer:
for j, image_name in enumerate(fold):
image_file = LiTS_Dir / image_name
print("\r>> Converting fold {}, {}/{}, {}/{}".format(
i + 1, j + 1, len(fold), counter, num_images),
end="")
seg_file = image_file.parent / image_file.name.replace(
"volume", "segmentation")
label_reader.read(seg_file)
bbox = array_kits.extract_region(label_reader.image(), align,
padding, min_bbox_shape)
label_reader.bbox = bbox.tolist()
tumor_value = np.max(label_reader.image())
indices = np.where(
np.max(label_reader.image(), axis=(1,
2)) == tumor_value)[0]
label_reader.indices = indices
tps = array_kits.find_tp(array_kits.merge_labels(
label_reader.image(), [0, 2]),
split=True)
# list of fps is sorted by z, y, x
for example in bbox_to_examples(label_reader, tps):
writer.write(example.SerializeToString())
counter += 1
print()
def check_tumor_hist(num=10,
xrng=(-200, 250),
bins=100,
yrng=(0, 0.02),
show=True,
save_path=None):
name = "volume-{}.nii".format(num)
image_path = misc.find_file(LiTS_ROOTS, name)
print(image_path)
mask_path = image_path.parent / image_path.name.replace(
"volume", "segmentation")
_, image = nii_kits.nii_reader(image_path)
_, mask = nii_kits.nii_reader(mask_path)
analysis_kits.compute_liver_tumor_hist(
image,
mask,
1,
2,
"{} - total".format(image_path.stem),
xrng=xrng,
bins=bins,
yrng=yrng,
show=show,
save_path=Path(save_path or "") / "0-total.png")
# For each tumor
disc = ndi.generate_binary_structure(3, connectivity=1)
labeled, num_obj = ndi.label(array_kits.merge_labels(mask, [0, 2]), disc)
mask = np.clip(mask, 0, 1) + labeled
for i in range(num_obj):
z, y, x = np.mean(np.array(np.where(mask == i + 2)),
axis=1).astype(np.int32)
analysis_kits.compute_liver_tumor_hist(
image,
mask,
1,
2 + i,
"{} - tumor {} xyz: ({}, {}, {})".format(image_path.stem, i + 1, x,
y, z),
xrng=xrng,
bins=bins,
yrng=yrng,
show=show,
save_path=Path(save_path or "") / "{}-tumor.png".format(i + 1))
def setUp(self):
data_path = Path(__file__).parent.parent / "data/LiTS/Samples"
_, self.image = mhd_kits.mhd_reader(data_path / "T001.mhd")
_, self.label = mhd_kits.mhd_reader(data_path / "T001_m.mhd")
self.image = array_kits.aug_window_width_level(self.image, 450, 50)
self.label = array_kits.merge_labels(self.label, [0, 255, 510])