Despite the name, this is a research training pipeline for 3D deep-learning segmentation models on liver disease ablation imaging (liver and ablation region masks). It sweeps hyperparameters and learning rates, trains DenseNet121-hybrid / U-Net style models in TensorFlow (parallel TF1 and TF2 script variants), and evaluates the results.
Dormant research code, last touched 2021; not maintained. It is not runnable standalone:
it imports local packages (Base_Deeplearning_Code, Deep_Learning) that are not included
in this repo, and data paths are hard-coded to internal drives.
Main.py,Main_TF2.py,Main_TF2_HNet.py— driver scripts; boolean flags step through the workflow: find best learning rate, plot LR curves, train ~200 epochs, tabulate results.Optimization/— LR range-finder (Find_Best_LR*.py) and plotting of LR/optimization results, logged via TensorBoard HParams.Return_Train_Validation_Generators*.py— TFRecord-based train/validation data generators and hyperparameter grids (layers, filters, max filters).Utils/— model builders (pretrained DenseNet121 with frozen encoder / trainable upsampling path), generators, path helpers, Excel metric export.Main_Evaluation*.py,Model_Test.py— model evaluation.Direct_Testing_From_Raystation/— scripts to export an examination and Liver/Ablation ROIs from RayStation and convert the MHD exports to NIfTI with SimpleITK, for testing models directly against treatment-planning-system data.
Python, TensorFlow 1.x/2.x, TensorBoard HParams, SimpleITK, pandas/openpyxl.