def __getitem__(self, idx):
"""
"""
dataset = H5Dataset(self.dataset_name, self.data_dir, mode="read")
exam = self.exams_df.iloc[int(idx)]
label = exam['label']
exam_id = exam['exam_id']
patient_id = exam['patient_id']
report = self._get_report(exam, dataset)
report = self.vocab.tokenize(report)
if self.max_len is not None and self.max_len < len(report):
report = report[:self.max_len]
report = self.vocab.wrap_sentence(report)
report, mask_labels = self._mask_inputs(report)
inputs = {"report": report}
targets = {"mlm": mask_labels, "abnorm": torch.tensor(label)}
info = {"exam_id": exam_id, "patient_id": patient_id}
return inputs, targets, info
def __getitem__(self, idx):
"""
"""
dataset = H5Dataset(self.dataset_name, self.data_dir, mode="read")
exam = self.exams_df.iloc[int(idx)]
label, exam_id, patient_id = exam['label'], exam['exam_id'], exam[
'patient_id']
report = self._get_report(exam, dataset)
images = self._get_images(exam,
dataset) if not self.skip_scans else None
targets = {}
# must perform scan_match first
if "scan_match" in self.task_configs:
args = self.task_configs["scan_match"]["args"]
report, labels = self.scan_match(exam_id, report, dataset, **args)
targets["scan_match"] = labels
# tokenize, trim if over max length, and wrap sentence
report = self.vocab.tokenize(report)
if self.max_len is not None and self.max_len < len(report):
report = report[:self.max_len]
report = self.vocab.wrap_sentence(report)
if "scan_mlm" in self.task_configs:
args = self.task_configs["scan_mlm"]["args"]
report, labels = self.scan_mlm(exam_id, report, **args)
targets["scan_mlm"] = labels
info = {"exam_id": exam_id, "patient_id": patient_id}
inputs = {"report": report, "scan": images}
return inputs, targets, info
def get_targets(self,
tasks=["abnormality_detection", "report_generation"]):
"""
"""
dataset = H5Dataset(self.dataset_name, self.data_dir, mode="read")
for idx, exam in self.exams_df.iterrows():
yield self._get_targets(exam, dataset, tasks=tasks)
def get_report(self, exam_id: str, split: str="valid"):
"""
"""
h5_dataset = H5Dataset(self.datasets[split].dataset_name,
self.datasets[split].data_dir,
mode="read")
report = h5_dataset.read_reports(exam_id)
return report
def __getitem__(self, idx):
"""
"""
dataset = H5Dataset(self.dataset_name, self.data_dir, mode="read")
exam = self.exams_df.iloc[int(idx)]
label = exam['label']
exam_id = exam['exam_id']
patient_id = exam['patient_id']
report = self._get_report(exam, dataset)
return label, report, {"exam_id": exam_id, "patient_id": patient_id}
def __getitem__(self, idx):
"""
"""
dataset = H5Dataset(self.dataset_name, self.data_dir, mode="read")
exam = self.exams_df.iloc[int(idx)]
label = exam['label']
exam_id = exam['exam_id']
patient_id = exam['patient_id']
inputs = self._get_images(exam, dataset)
targets = torch.tensor(label)
del dataset
return inputs, targets, {"exam_id": exam_id, "patient_id": patient_id}
def __getitem__(self, idx):
"""
"""
dataset = H5Dataset(self.dataset_name, self.data_dir, mode="read")
exam = self.exams_df.loc[int(idx)]
label = exam['label']
exam_id = exam['exam_id']
patient_id = exam['patient_id']
images = self._get_images(exam, dataset)
targets = self._get_targets(exam)
info = {"exam_id": exam_id, "patient_id": patient_id, "idx": idx}
return images, targets, info
def get_targets(self, tasks=[]):
"""
"""
dataset = H5Dataset(self.dataset_name, self.data_dir, mode="read")
for idx, exam in self.exams_df.iterrows():
yield self._get_targets(exam, dataset, tasks=tasks)
def show_saliency_slice(
exam_id: str,
image_slice: int,
show_pet: bool = False,
saliency: np.array = None,
data_dir: str = "/data/fdg-pet-ct",
dataset_name: str = "pet_ct_dataset",
ct_hu_range: tuple = 0,
pet_percentile: float = 0,
saliency_percentile: float = 0,
process_dir: str = None,
):
"""
### <<EXPERIMENT SPEC>> ###
# axillary
#experiment_dir = "experiments/manuscript/_seed/single_task/08-01_12-42_st-pretrain_mt_full_26_e/candidates/exp_15"
#cervical
#experiment_dir = "experiments/manuscript/_seed/single_task/08-01_12-42_st-pretrain_mt_full_26_a/candidates/exp_19"
#inguinal
experiment_dir = "experiments/manuscript/_seed/single_task/08-01_12-42_st-pretrain_mt_full_26_a/candidates/exp_5"
"""
dataset = H5Dataset(dataset_name, data_dir, mode="read")
ct = np.array(dataset.read_images(exam_id, "CT Images"))[image_slice]
pet = np.array(dataset.read_images(exam_id, "PET_BODY_CTAC"))[image_slice]
plt.figure(figsize=(10, 10))
print(ct.max())
print(ct.min())
# plot ct
plt.imshow(
plt.cm.bone(
plt.Normalize(vmin=ct_hu_range[0], vmax=ct_hu_range[1], clip=True,)(ct)
)
)
# plot pet
# set alpha based on pet instensity
if show_pet:
pet = cv2.resize(pet, ct.shape)
alphas = plt.Normalize(clip=True)(np.abs(pet))
pet = plt.cm.plasma(
plt.Normalize(vmin=np.percentile(pet, pet_percentile), clip=True)(pet)
)
pet[..., -1] = alphas
plt.imshow(pet)
# plot saliency
if saliency is not None:
saliency = saliency.max(axis=4, keepdims=False)
saliency = saliency - saliency.min()
saliency /= saliency.max()
saliency = saliency[0, image_slice]
saliency = cv2.resize(saliency, ct.shape)
alphas = plt.Normalize(
vmin=np.percentile(saliency, saliency_percentile), clip=True
)(np.abs(saliency))
saliency = plt.cm.viridis(plt.Normalize()(saliency))
saliency[..., -1] = alphas
plt.imshow(saliency)
if process_dir is not None:
plt.savefig(os.path.join(process_dir, "scan.png"))