Deep Learning in Medicine resources & Study Group

Part 1 (2019)
43

Hi,
Iā€™m an emergency physician and researcher in Northern California, hoping to apply ML tools to predicting sepsis. But regardless of your particular health+ML interest, if you ever need to chat to a physician about domain knowledge, just get in touch!

5 Likes
44
1 Like
45

Hello!

Michael Woodburn here, now working as a Doctor in Australia. Hoping to contribute to the ML field by creating the best dataset possible from our radiological images. I have access to 100,000 CT Abdominal Images, their metadata (name, age, etc), and freetext Radiologist-written reports. To make this dataset useful I am going to label the whole thing via various methods (automated, and human).

All reports look something like this:

ā€œā€"
CLINICAL HISTORY: ABDOMINAL PAIN.
TECHNIQUE: Multiple axial CT images were obtained through the abdomen and pelvis after
administration of oral contrast material only.
COMMENTS:
There is evidence of diffuse hepatic hypoattenuation compatible with fatty infiltration. There is no
intra or extrahepatic biliary ductal dilatation. The patient is status post cholecystectomy. The spleen
is normal. The pancreas is of normal contour and attenuation characteristics. There is no evidence of
adrenal mass.
Moderate sized fat containing supraumbilical hernia is present.
The kidneys are normal in size, shape and configuration. No renal or ureteral calculi are identified.
There is no hydroureter or hydronephrosis.
There is no evidence for appendicitis. Several fluid-filled loops of small bowel are present compatible
with mild enteritis. There is no bowel wall thickening. No evidence for small or large bowel
obstruction. There is no evidence of abdominal ascites or lymphadenopathy.
There is no evidence of intrinsic or extrinsic bladder mass. There is no pelvic ascites or
lymphadenopathy.
The uterus and ovaries are grossly unremarkable.
Images of the lung bases show no evidence of pleural or parenchymal mass. There are no pleural
effusions. Scarring is present in the right middle lobe and lingula as well as both lung bases.
The bony structures are free of lytic or blastic lesions. Multilevel degenerative changes are seen
involving the thoracolumbar spine.
Scattered calcifications are seen involving the aorta and major branches compatible with
atherosclerosis. IMPRESSION:
ABDOMEN:

  1. Several fluid-filled loops of small bowel are present compatible with mild enteritis.
  2. Fatty liver.
  3. Status post cholecystectomy.
  4. Fat containing umbilical hernia.
    PELVIS:
  5. No evidence of diverticulitis or acute inflammatory process in the pelvis.
    ā€œā€"

And we have some idea of the specific labels we will derive from these reports. For example:

  1. Pathologies (diseases, e.g appendicitis)
  2. Radiological features (e.g fat stranding)
  3. Report written by trainee vs senior
  4. Follow up study required
  5. Scan of sufficient quality to report

My big question to you all is
Can anyone think of any other interesting labels/targets we should be looking for?
Something a little left of field perhaps?

Open to any and all suggestions. Looking forward to making this dataset available to interested ML enthusiasts in future.

2 Likes