Artificial

Intelligence in

Medicine

AIM1 - Data Harmonization

​

Joint Analysis Pipeline for Data Harmonization in MRI and PET

​

A. Retico (PI)

AIM, AIM1 Meeting - Febr 17, 2021 https://agenda.infn.it/event/25883/

AIM1.T1 AIM1.T3

​

AIM, CSN5, 2019-2021

AIM1 - multi-site data harmonization

data gathered by different sites and/or acquisition systems carries local “fingerprint”, often to the detriment of the much more subtle information of interest.

​

this problem is akin to the management �of systematic errors

​

typical application cases: MRI, RX, PET, NPSY tests

Autism Brain Imaging Data Exchange

AIM, CSN5, 2019-2021

Data normalization strategy already implemented in AIM

• Intra-subject Normalization (Pisa, Ferrari et al. 2020, AIIM 108, 101926. https://doi.org/10.1016/j.artmed.2020.101926):
• volumetric features are divided by the total intracranial volume
• cortical surfaces are divided by the area of the total white matter
• cortical thicknesses are divided by the mean cortical thickness across the entire brain.
• Adjust for know batches using COMBAT [1,2] (Bologna)
• Adjust for know batches using COMBAT [1,2] (Bari, Lombardi A, Amoroso N., Diacono D., Monaco A., Tangaro S., Bellotti R. Extensive Evaluation of Morphological Statistical Harmonization for Brain Age Prediction, Brain Sciences, 2020, 10(6), 364)

​

​

[1] Johnson WE, Li C, Rabinovic A. Adjusting batch effects in microarray expression data using empirical Bayes methods. Biostatistics. 2007 Jan 1;8(1):118-27.

[2] Fortin, J.P.; Parker, D.; Tunç, B.; Watanabe, T.; Elliott, M.A.; Ruparel, K.; Roalf, D.R.; Satterthwaite, T.D.; Gur, R.C.; Gur, R.E.; et al. Harmonization of multi-site diffusion tensor imaging data. Neuroimage 2017, 161, 149–170.

​

​

3

AIM, CSN5, 2019-2021

Recently…

Dataset of 10477 typical subjects [3-96 years]

​

ComBat-GAM

Not limited to linear model for age trends, but introduces Generalized Additive Model (GAM)

4

non-linear function of age, sex, TIV

location

scale

AIM, CSN5, 2019-2021

Age trends for selected ROI (Pomponio et al. 2020)

5

AIM, CSN5, 2019-2021

Final considerations from Pomponio et al. 2020

• The authors made publicly-available:
• visualization tool we provide as a product of the LIFESPAN dataset (https://rpomponio.shinyapps.io/neuro_lifespan/)
• a package that enables users to apply ComBat-GAM on their own datasets (https://github.com/rpomponio/neuroHarmonize)

​

6

AIM, CSN5, 2019-2021

Proposed pipeline for AIM1 joint analysis on Data Harmonization

• Milestone 2021 related to AIM1
• AIM.1: Valutazione dell’impatto delle diverse strategie implementate per l’armonizzazione dei dati e identificazione delle strategie ottimali rispettivamente per studi MRI/Mammografici/PET multicentrici

​

• We are probably ready to define a common strategy for MRI and PET data harmonization. Breast imaging will be handled in a separate work.

​

7

AIM, CSN5, 2019-2021

Proposed pipeline for AIM1 joint analysis on Data Harmonization

Paper outline (“Impact of data harmonization on ML performance in multicenter studies: MRI, fMRI and PET applications”):

• Introduction on the need for data harmonization (DH) in multicenter studies
• Overview of most used techniques (focusing on MRI, fMRI and PET)
• Focus on the impact of DH in machine learning (ML) applications (limiting to case-control classification)
• Each clinical question deserves a specific harmonization strategy
• (harmonizing raw data vs. harmonizing extracted features; which DH strategy?)
• Case studies:
• selection of available/accessible “large” datasets
• selection of clinical questions that can be answered with ML (classification/regression)
• evaluation of the effect of DH on performance (choice of appropriate evaluation metrics)
• Discussion and definition of “general guidelines” for DH

​

8

AIM, CSN5, 2019-2021