Jordi

[Jordi @jdbardia]

I'm guessing this model trained on the 3D data would be applicable for segmenting 2D data, as well?

[Jordi @jdbardia]

#SMILEJournalClub I read a paper about a self-supervised transformer model applied to 3D medical image data that got impressive results in a segmentation task. This seems like it would be incredibly useful for dealing with medical data when you are not a medical professional!

[Jordi @jdbardia]

learning. I wonder what the effect of switching the initial training data set would be on the transfer learning outcomes. Makes me wonder further on what images a "Medical ImageNet" would consist of, for training these sorts of models.

[Jordi @jdbardia]

Today's Journal Club paper is on "Models Genesis", a self-supervised machine learning model that can learn from different imaging modalities and from different imaging perspectives. For classification and segmentation tasks, it manages to outperform... #SMILEJournalClub

[Jordi @jdbardia]

even models that were fully supervised during training. Impressive! From my understanding of the paper, the authors trained the model using chest CT scans and were able to perform the aforementioned tasks on other parts of the body and attain good results via transfer...

[Jordi @jdbardia]

on, based on the pathways from the visual cortex to the amygdala. The amygdala network has a "middle road" too, so I wonder what the result would be if more pathways are included to the paper's model.

[Jordi @jdbardia]

This week's journal club paper in on a machine learning model that uses parallel pathways inspired by the dorsal and ventral pathways in the brain responsible for visual recognition and movement, respectively. This is similar to a model that I've been working...#SMILEJournalClub

[Jordi @jdbardia]

CNNs should always be used over 2D CNNs, or is there some sort of advantage to using 2D CNNs, possibly in conjunction with the 3D CNNs?

[Jordi @jdbardia]

This week's journal club was pretty surprising to me--it points out the flaws of the methodology of several papers on Alzheimer's classification, which manage to pass through peer review and be published in journals. This could potentially cause harm... #SMILEJournalClub

[Jordi @jdbardia]

down the line, so it's good they develop a reproducible method for proper evaluation. On a separate note, the paper mentions 3D CNNs consistently performing better than 2D CNNs in terms of classification accuracy. Does this suggest that for AD classification, 3D...

[Jordi @jdbardia]

Also interesting to see how choosing random weight seeds has a similar trend in representational consistency to using different images. I wonder how different the images used in each group were, and if the difference will affect the rep. consistency.

[Jordi @jdbardia]

Pretty cool paper. Coincidentally, it's somewhat relevant to what my group is working on right now. I wonder how the variance between models will be affected when some of the weights are kept the same, and if local recurrent processing would reduce it. #SMILEJournalClub

[Jordi @jdbardia]

as well as a whole list of 1057 SNPs that were found to be associated under a certain p-value*

[Jordi @jdbardia]

Interesting to see how CNNs are applied with other data to make predictions in medical research. The inclusion of SNP features combined with the image analysis seems to have a miniscule(?) effect on the prediction rate as can be seen by the AUC values...#SMILEJournalClub

[Jordi @jdbardia]

these SNPs have intercorrelations or not, which may have lowered the AUC values of the two analyses. Some sort of selection process (like a forward/backward stepwise procedure) could maybe be used to get a group of SNPs that predict AMD with a higher accuracy.