Closed mxochicale closed 2 years ago
as discussed with you last week, I think the relevant background to 4CV is the segment when the doctor started looking for 4 chamber view and when they stop the acquisition in the apex of the heart
_Originally posted by @huynhatd13 in https://github.com/vital-ultrasound/echocardiography/pull/36#discussion_r816743827_
Considering our previous meetings and input from Nhat, I have sketched the following diagram that might help to distil our thinking on which classes might be relevant and how to select which can be hard and soft labels.
Not too much bandwidth to address this one, so I close it. Feel free to open it.
🚀 Feature
In our today's weekly meeting we touched on the use of soft-labels for 4 Chamber View echos and how class balance will play an important role for the case of 10 minutes echocardiography videos. One potential avenue is the use hard labels for A4c no occlusions, A4c occluded LA and A4c occluded LV Zhang et al. in Circulation. However, according to Nhat, those occlusion labels in the context of the ICU might not be considered but perhaps the use of good and bad views for labels (considering, perhaps, OxAxis, Clarity, DepthGain, Fshorten features as in 2021-labs-miua). Perhaps we might like to made use of something like "The soft label format [Probability of Benign, Probability of Malignancy]" as done in Cao, Z., Yang, G., Li, X. et al. Multitask Classification Method Based on Label Correction for Breast Tumor Ultrasound Images. Neural Process Lett 53, 1453–1468 (2021) but require further discussions.
One potential approach is to sketch a pipeline for something like "[Probability of Entering 4CV, Probability of being in 4CV]" and one protocol for creating those soft labels. But we need to be careful:
"POSITIVES OF SOFT_LABELS: By using real numbers instead of single bits, soft-labels provide to the learning algorithm extra information that often re- duces the number of instances required to train a model [15], while improving the performance during inference [5, 6, 18]." CHALLENGES OF SOFT_LABELS "The main challenge in using soft labels is their proper computation. One complication is that human experts struggle to give reliable and consistent estimates of the probabilities. One effective way of reducing this problem is to group the probabilities into bins [32]. However, this still relies on human estimates" 2021-Vega-in-arXiv
Motivation
The current implementations are just cosidered hard labels for background and 4CV to which the extension for soft-labels might be desirable.
Pitch
Alternatives
Additional context
To have a look: Yu Hung et al. 2020 in arXiv