Deep-learning approach for tissue classification using acoustic waves during ablation with an Er:YAG laser

This is the code to the paper Deep-learning approach for tissue classification using acoustic waves during ablation with an Er:YAG laser (Updated) (Arxiv 2024).

ATTENTION

This This manuscript is currently undergoing revision due to issues identified with the data we received. It was found that some of the training, testing, and validation datasets were scaled relative to each other. This scaling inadvertently resulted in an unrealistically high accuracy rate of 100% for our neural network.

Upon recognizing this data-related issue, we conducted a thorough reevaluation by repeating the experiments. These revised measurements revealed a significantly lower accuracy rate, falling below 40% – a figure markedly lower than what was initially anticipated.

In response to these findings, we have extensively revised and enhanced our network architecture. The improvements made have led to a substantially better-performing model. Our optimized network now achieves a significantly higher accuracy rate, exceeding 75%. This notable enhancement marks a considerable improvement in both the performance and reliability of our model.

Branches Description

• branch old: initial code of the paper.
  • issues with the data (some of the data were scaled relative to each other)
  • unrealistic accuracy of 100% of the neural network
  • published initially in IEEE Access 2021
• branch old_with_new_data: repetion of the experiment. And used new data to retrain the neural network
  • accuracy was under 40%
  • is described in correction of paper
• branch main: used the new data and in depth analisis and fine tuning of the neural network
  • best performing network: exceeds 75% accuracy
  • detailed description can be found on the corrected paper

Branch main

This is based on the new measurments from 11.2022. It uses new network to get better accuracy. This should be used as comparison for future work.

• Folder describe the different networks
  • ANN simple fully conected network with the first three PCA of the fft as an input
  • FcNN_pca is a fully connected network with PCA of the fft as an input
  • FcNN_fft is a fully connected network with fft as an input of the neural network
  • CNN_time is a convolutional neural network with the time dependent data as input
  • CNN_FcNN a combination of CNN and FC which uses time and fft as input
• subfolder are in the form of *{0,1,2,3,4}_trans1_batchnorm_Multiinput{_1,_2,_5,_10}
  • {0,1,2,3,4} determain the different crossvalidations
  • {_1,_2,_5,_10} the number of successive acoustic waves
• data are divided into 10 subsets represented by the numbers `1,...,9,0'
• the data are pared together as follow: 16,27,38,49,50
• these are used for training, testing, and validation accordingly to the files in parameter
• adapt the path to the desired folder where the data are saved
• test_eryag_*_length_all.py is used to test the performance of the network
• train_model_0.py is used to train the network
• train_optuna_0.py is used to do a hyperparameter search with optuna
• activityMap.m plots the activation map previously computed
  • CNN_time_*_trans1_batchnorm_MultiInput_1 can be computed with the script activationMap_GradCam.py

Data

Please contact Prof. Dr. Philippe C. Cattin ([email protected]) to access the data

The data can be found on lakeFS: https://dbe-lakefs.dbe.unibas.ch/repositories/tissue-classification-acoustic

• branch old: some of the data were scaled relative to each other.
• See compareData.m to see an examle of the issues of the data for the measurment with the Er:YAG and Nd:YAG data.
• The initial publication has issues with the high accuracy.
• code for inital publication can be found in the branch old on GitHub and GitLab
• branch main: new measurment done on the 22.-24.11.2022.
  • saveAsMatrix.m transforms the data into a matrix and saves them in the folder matrix_all
  • settings of laser and the measurment of the acoustic signal, see details in Paper corrected paper
• branch old_with_new_data: the data und the folder matrix_GradCam_old uses the new measursments (11.2022) with the initial code from the inital paper
  • code can be found in the branch old_with_new_data on GitHub and GitLab
  • in the folder matrix_GradCam_old is the file cut_to_size_preprocess.m which transforms the matrix matrix_all in the desired format
• branch remake: the data can be found under matrix_pre_3000
  • code can be found in the main branch on GitHub and GitLab
  • in the folder matrix_pre_3000 is the file preProcessing_pre_3000.m which transforms the matrix matrix_all in the desired format