LE Artificial intelligence for medical image analysis
Learning objectives:
General introduction to AI – how does it work
- Classifiers powered by deep learning, trained on human labels
- Train/val/test set
How can AI help in medical image analysis?
- Speed up case reading
- Go beyond the reading performance of humans
Why is (good) AI difficult to develop?
- Data quality and quantity
- Establishing a reference standard
3 stages of AI
- Narrow AI, assist with or take over tasks (google maps)
- General AI, takes knowledge from domain, transfers to other domain (autonome keuze
maken)
- Super AI, machines that are an order of magnitude smarter than humans (matrix)
Machine learning
- Decide on features that describe the difference between the categories you want to classify
- Build a model that learns how to discriminate between the categories using the features
Drawbacks of manual feature extraction
- Laborious (arbeidsintensief)
- Prone to error
- Not always effective
Deep learning: let the model learn from the data itself
- Do not prescribe what model should learn
- More neural network layers, so features are getting more abstract and predict the outcome
better
Hyperparameters are used to improve the algorithm (draaiknoppen van algoritme)
Applications of DL in medical imaging
- Image registration
- Computer-aided diagnostics
- Workload reduction
Image registration: aligning multiple perspectives
- Tissues are not static
o Breathing intrascan movement
o Bladder/bowel interscan movement
o Patient (intra/interscan) gross movement
- To align information from different time points, you need image registration
Computer aided diagnostics (CAD)
- Automatic interpretation of medical images, either for decision support for the physician or
replacing the physician entirely
o Decision-support via finding pathologies/abnormalities that are in the clinical
protocol
, o Replacement: automating every part of a clinical guideline
AI had multiple unsolved problems
- Fundamental, unsolved problem in AI:
o Generalization to different medical centres
- Practical problems:
o Interobserver variability of annotators
o Dataset size
- Problems cause by the designer of the AI:
o Bias introduced by poor methodological design
LE Interactive lecture
Case 1:
Tumor growth through muscularis propria into serosa
15/15 lymph nodes without malignant cells
- T3
- N0
- M0
- (T4 = when goes to distant organs)
- Additional test immunohistochemistry, mismatch repair
Interaction of the mismatch repair proteins
- MSH2/MSH6
- MLH1/PMS2
Lynch syndrome
- Often mutation in mismatch repair system
- Check with DNA sequencing in blood whether there is defect in for instance PMS2
Case 2:
Sentinal node = first lymph node that causes the tumour
Case 3:
- Inflammation leads to immune cells and therefore oxygen species are release and leads to
damage of the bowel leading to mutations and tumours.
- Stroma cells are in connective tissue, these are coloured brown and could look like they are in
the tumour tissue, but then still the IHC is negative
- Look at methylation through PCR (hypermethylation when there is mismatch defect)
o Use PCR and enzymes that are not cutting in methylated DNA
o Peak means there is DNA
o When using enzyme, areas that are methylated are not shown in PCR
o When there is tumour, enzyme can not digest them and are still shown in PCR
When patient has metastatic disease and mismatch repair deficiency, use checkpoint inhibitor
Case 4:
Positive stool test = test for blood in poo (gastrointestinal issues) in population screening
- Test for blood with antibodies
- Mucinous adenocarcinoma = arise from epithelial cells, some cells produce mucus for protective
barrier, sometimes tumour cells produce also lot of slime, have more often MSI
- Proximal colon easier to remove tumours because there is more space than in distal colon
Learning objectives:
General introduction to AI – how does it work
- Classifiers powered by deep learning, trained on human labels
- Train/val/test set
How can AI help in medical image analysis?
- Speed up case reading
- Go beyond the reading performance of humans
Why is (good) AI difficult to develop?
- Data quality and quantity
- Establishing a reference standard
3 stages of AI
- Narrow AI, assist with or take over tasks (google maps)
- General AI, takes knowledge from domain, transfers to other domain (autonome keuze
maken)
- Super AI, machines that are an order of magnitude smarter than humans (matrix)
Machine learning
- Decide on features that describe the difference between the categories you want to classify
- Build a model that learns how to discriminate between the categories using the features
Drawbacks of manual feature extraction
- Laborious (arbeidsintensief)
- Prone to error
- Not always effective
Deep learning: let the model learn from the data itself
- Do not prescribe what model should learn
- More neural network layers, so features are getting more abstract and predict the outcome
better
Hyperparameters are used to improve the algorithm (draaiknoppen van algoritme)
Applications of DL in medical imaging
- Image registration
- Computer-aided diagnostics
- Workload reduction
Image registration: aligning multiple perspectives
- Tissues are not static
o Breathing intrascan movement
o Bladder/bowel interscan movement
o Patient (intra/interscan) gross movement
- To align information from different time points, you need image registration
Computer aided diagnostics (CAD)
- Automatic interpretation of medical images, either for decision support for the physician or
replacing the physician entirely
o Decision-support via finding pathologies/abnormalities that are in the clinical
protocol
, o Replacement: automating every part of a clinical guideline
AI had multiple unsolved problems
- Fundamental, unsolved problem in AI:
o Generalization to different medical centres
- Practical problems:
o Interobserver variability of annotators
o Dataset size
- Problems cause by the designer of the AI:
o Bias introduced by poor methodological design
LE Interactive lecture
Case 1:
Tumor growth through muscularis propria into serosa
15/15 lymph nodes without malignant cells
- T3
- N0
- M0
- (T4 = when goes to distant organs)
- Additional test immunohistochemistry, mismatch repair
Interaction of the mismatch repair proteins
- MSH2/MSH6
- MLH1/PMS2
Lynch syndrome
- Often mutation in mismatch repair system
- Check with DNA sequencing in blood whether there is defect in for instance PMS2
Case 2:
Sentinal node = first lymph node that causes the tumour
Case 3:
- Inflammation leads to immune cells and therefore oxygen species are release and leads to
damage of the bowel leading to mutations and tumours.
- Stroma cells are in connective tissue, these are coloured brown and could look like they are in
the tumour tissue, but then still the IHC is negative
- Look at methylation through PCR (hypermethylation when there is mismatch defect)
o Use PCR and enzymes that are not cutting in methylated DNA
o Peak means there is DNA
o When using enzyme, areas that are methylated are not shown in PCR
o When there is tumour, enzyme can not digest them and are still shown in PCR
When patient has metastatic disease and mismatch repair deficiency, use checkpoint inhibitor
Case 4:
Positive stool test = test for blood in poo (gastrointestinal issues) in population screening
- Test for blood with antibodies
- Mucinous adenocarcinoma = arise from epithelial cells, some cells produce mucus for protective
barrier, sometimes tumour cells produce also lot of slime, have more often MSI
- Proximal colon easier to remove tumours because there is more space than in distal colon
