Med-Real2Sim: Non-Invasive Medical Digital Twins using Physics-Informed Self-Supervised Learning
Authors: Keying Kuang, Frances Dean, Jack B. Jedlicki, David Ouyang, Anthony Philippakis, David Sontag, Ahmed M. Alaa
NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We apply our method to identify digital twins of cardiac hemodynamics using noninvasive echocardiogram videos, and demonstrate its utility in unsupervised disease detection and in-silico clinical trials. We train a 3D-CNN implemented using Pytorch to output a subset of the model parameters θi. Our approach, Med-Real2Sim, achieves a mean absolute error (MAE) of 6.81% for CAMUS and 5.40% for Echo Net in predicting EF (Table 1) |
| Researcher Affiliation | Collaboration | Keying Kuang UC Berkeley Frances Dean UC Berkeley & UCSF Jack B. Jedlicki University of Barcelona David Ouyang Cedars Sinai Anthony Philippakis Google Ventures David Sontag MIT Ahmed Alaa UC Berkeley & UCSF |
| Pseudocode | No | The paper describes methods using text and mathematical equations but does not include explicit pseudocode or algorithm blocks. |
| Open Source Code | Yes | Justification: We use publicly available data and our code is already publicly available on Git Hub. |
| Open Datasets | Yes | We test our physics-informed SSL (Med-Real2Sim) approach using two echocardiography datasets: Echo Net and CAMUS. The CAMUS dataset (44) consists of 500 fully annotated cardiac ultrasound videos... The Echo Net dataset (53) comprises 10,030 apical-4-chamber echocardiography videos from routine clinical care at Stanford University Hospital |
| Dataset Splits | Yes | The CAMUS dataset was split into 450 training samples and 50 validation and testing samples. The Echo Net dataset was partitioned into 7,466 training, 1,288 validation, and 1,276 testing samples. |
| Hardware Specification | Yes | Models were trained on CPUs, each node featuring 24 cores (12 physical cores with hyper-threading) and 128 GB of RAM. |
| Software Dependencies | Yes | We train a 3D-CNN implemented using Pytorch to output a subset of the model parameters θi. PINNs were implemented in Tensorflow (16). Neural ODEs were implemented in Pytorch using torchdiffeq (11). The network bϕM is pretrained by generating 3,840 synthetic data points ... using a numerical ODE solver (73). [Reference 73: Sci Py 1.0: Fundamental Algorithms for Scientific Computing in Python. Nature Methods, 17:261 272, 2020.] |
| Experiment Setup | Yes | The Echo Net model, achieving a Mean Absolute Error (MAE) of 5.40%, was trained for 3 hours over 13 epochs with a batch size of 100 and a learning rate of 0.001. In comparison, the CAMUS model, which reached an MAE of 6.81%, required 2 hours of training over 110 epochs with a batch size of 50 and a learning rate of 0.005. ... We evaluated learning rates of 0.001, 0.005, and 0.0005, batch sizes of 50 and 100, and the number of epochs set to 300 and 500. We used the adam optimizer for training. |