DeepLearning CalciumDetection

Coronary artery calcifications are often overlooked in routine clinical practice, and even for an experienced radiologist it is difficult to detect calcium from an X-ray image by eye compared to a CT scan. Chest x-rays are obtained much more commonly than CT scans at a much lower cost; for this reason an automatic radiographic image analysis system could be of fundamental importance for the early detection of calcium. Through this work we aim to build a process, based on convolutional neural networks, capable of predicting the presence or absence of coronary calcium from chest radiographs, evaluating the performance in terms of accuracy obtained.

Data

For the correct functioning of the code shown in this repo it is necessary to work with the dataset content in the folder /home/fiodice/project/dataset. The code in script_clean_data/clean_dataset.py is a script for cleaning the dataset, but many labels of the dicom files are not populated. The script removes most of the useless files, but a manual cleanup was performed for the rest.

Training

Two approaches were performed for solving the task, for both approaches a cross validation with 5 fold was performed.

• In train_cac_classifier.py the classification approach is implemented
• In train_cac_regressor.py the regression approach is implemented

For the files you can specify the following parameters:

• epochs : num. of epochs (default 50)
• lr : learning rate default (3e-4) for classification and (1e-3) for regression
• arch : encoder architecture (densenet121 or resnet18 or efficientNet)
• viz : save images of metrics and losses
• save : save model
• wd : weight decay value (default 1e-4)
• batchsize : samples for batch
• momentum : momentum value (default 0.9)
• kfold : folds for cross-validation (default 5)

Result

The best models resulting with the relevant training details, could be found in the result models_cac_pt. More details on the methodology and quality of the results can be found in the thesis of Francesco Iodice followed by Professor Marco Grangetto and the co-supervisors Alberto Presta and Carlo Alberto Barbano.