Keywords:
Artificial Intelligence, CT, Computer Applications-Detection, diagnosis, Segmentation, Staging, Cancer
Authors:
A. Corsi, F. Belmans, F. Blistein, S. Ahmed Ali, W. Vos, S. Sironi, M. Occhipinti
DOI:
10.26044/ecr2023/C-20015
Methods and materials
For the training and validation, we used the publicly available CT Lymph Nodes dataset [3] (90 CECTs with manual segmentations performed by radiologists). The authors revised each mask to add every mediastinal lymph node (minor axis > 10 mm) that was not segmented initially and the ones showing distinctive pathological features (minor axis range: 5-10 mm) (See example in Fig.1).
Particular attention was used to be as adherent as possible to lymph nodes borders. Lymph nodes with the tendency to merge into conglomerates were preferentially segmented as individual ones as long as they were distinguishable from each other. Otherwise, they were segmented as a conglomerate. The final number of mediastinal lymph nodes segmented was 454 in the original dataset and 1162 after revision and corrections. The proposed architecture is based on a U-Net Convolutional Neural Network with a ResNeXt-50 backbone for feature extraction and deep supervision at the output layers. We optimized the model’s weights for Tversky loss using the Adam optimizer (learning rate: 0.0001). The same model was trained on both datasets using 5-fold cross-validation. DICE similarity score (DSC) and relative absolute volume difference (RAVD) were used for the model’s performance evaluation.