AnDiChallenge2

This is the inference code of team HNU for the single trajectory task in AnDi Challenge2.

List of team members: Zihan Huang, Xiang Qu

Affiliations: School of Physics and Electronics, Hunan University, Changsha 410082, China

Email of the Team leader: [email protected]

Name of the method: WADNet + U-AnDi

Brief description of the method:
Our method for the single trajectory task in track 2 combines the capabilities of two networks from our previous articles: WADNet [1] and U-AnDi [2]. Here, WADNet is a network dealing with trajectory classification and regression task, while U-AnDi is a segmentation model making point-wise preditctions. Details of our method are listed as follows:

1. Initial classification with WADNet:
   The first step involves using WADNet to classify the model of each experiment (exp) as SSM, MSM, QTM, or TCM. In particular, for the DIM model, visual inspection suffices, and it is not processed through WADNet.

2. Traing WADNet and U-AnDi:
   For each classified model, independent models tailored to the trajectories are trained using both WADNet and U-AnDi:
   a) Classification model (WADNet): Identifies if there are changepoints within the trajectory.
   b) Regression model (WADNet): If no changepoints are detected, this model estimates the diffusion exponent (alpha) and the diffusion coefficient (K) directly.
   c) Segmentation model based on alpha (U-AnDi): Outputs the alpha for each point along the trajectory.
   d) Segmentation model based on K (U-AnDi): Outputs the K for each point along the trajectory.

3. Changepoint detection and parameter estimation:
   For trajectories categorized under SSM, MSM, QTM, or TCM, the following steps are employed:
   a) Determine the presence of changepoints using the classification model (2.a).
   b) If no changepoints are present, directly output alpha and K using the regression model (2.b).
   c) If changepoints exist, derive alpha and K for each point along the trajectory using models (2.c) and (2.d).
   d) Apply the post-processing technique in U-AnDi [2] to determine changepoints based on the variations in alpha.
   e) Segment the trajectory based on these changepoints, and calculate the average values of alpha and K from step (3.c) for each segment to predict segment properties.

4. Special case for DIM:
   Given the interactive nature of dimerization involving two trajectories, a brute-force matching approach is adopted for the DIM exps:
   a) Changepoints are identified by examining the distance between points of two trajectories in the same frame to confirm dimerization occurrences.
   b) Trajectories are then segmented based on these changepoints, and the alpha and K of each segment (length >= 20) are estimated using the regression model (2.b).
   c) For very short segments (length < 20), we use the average values of alpha and K from the entire experiment as their predictive values.

References:
[1] D. Li, Q. Yao, Z. Huang, WaveNet-Based Deep Neural Networks for the Characterization of Anomalous Diffusion (WADNet). J. Phys. A: Math. Theor. 2021, 54, 404003.
[2] X. Qu, Y. Hu, W. Cai, Y. Xu, H. Ke, G. Zhu, Z. Huang, Semantic Segmentation of Anomalous Diffusion Using Deep Convolutional Networks. Phys. Rev. Research 2024, 6, 013054.