braf mutation prediction from metastatic melanoma...
TRANSCRIPT
DIMED – University of Padova
BRAF mutation prediction from metastatic melanoma cytological smear
using deep learning: an ongoing project
Nicolè Lorenzo1, Rocco Cappellesso 1, Samir Suweis 2, Filippo Cappello 1, Kathrin Ludwig1 and Fassina Ambrogio1
1Department of Medicine, Surgical Pathology & Cytopathology Unit, University of Padova, Padova, Italy.2 Physics and Astronomy Department, University of Padova, Padova, Italy.
Background - Malignant Melanoma -
Melanoma in Veneto (2015)
1318/year (696M;622F)
Source: Veneto Tumor Registry -- https://www.registrotumoriveneto.it/en/
Molecular AnalysisTherapyFollow-up
Redirect to a Third level Structure
In our cytopathology service we detect about 100/year of metastatic melanomas by FNAB
Aim
Set up of a deep learning computational classifier able to classify BRAF mutated melanoma vs wild
type melanoma in cytologic smears.
Deep learning: …Dark side of pathology…
Materials & Methods –Methodology –
Cases selections Pathologist review
Patches extraction
Class attribution
CNN training
Retrain CNN after pathologist review and correct predicted results
Trained CNNs
Trained CNNs
Testing CohortCytological/molecular
confirmed Classified cases
BRAFm
FIRST STEP
SECOND STEP
• BRAFm• BRAFwt
Misclassification
Correct classification
STEP 3
STEP 2
BRAFwt
Trained CNNs
New cases
Concordance
Predicted mutation/WT
THIRD STEP
Confirmed mutation/WT
Conventional Workflow
Materials & Methods
Cases retrievaland review
Digitalization
273 cases (2016-2018)
901 Smears
111 MMG Smearsincluded
(94 cases: 64wt/30m)
Virtual Slides(Maximum magnification 40x)
Smear segmentation
Tiling: 100x100um tilesAbout 50.000 tiles/slide
About 6000000 tiles automatically exportedas single TIFF files
80% of tiles >>> training 20% >>> validation
Informative Non Informative
Nuclear segmentation
Not malignant
Removed from input files by post segmentation analysis
Input for CNN training
Qupath: Openslides –Smear segmentation -Tiling
Python: Tilies Screening –Nuclear segmentation –Post processing
Docker: Deep learningEnvironment
Tensorflow: Classification
Youtube tutorial: learn all above