Keywords:
Performed at one institution, Not applicable, Metastases, Lymphoma, Cancer, Computer Applications-General, CT, Lymph nodes, Artificial Intelligence, Abdomen, Artificial Intelligence and Machine Learning
Authors:
T. Sakinis, H. B. Jenssen; Oslo/NO
DOI:
10.26044/ecr2020/C-14337
Methods and materials
In our experiments, we have found that supplying size information to the CNN model helps to improve the segmentation accuracy. Therefore, we created a two-model approach based on DG.
Two-model approach
Model 1 is used to create a “rough” segmentation of the lymph node of interest. It has been trained with raw data and a fixed-sized mouse pointer as input. The output of this model is used to determine a size estimation that is close to the ground truth. We display the size estimation in the form of a variably sized mouse pointer (which is marked as a green dot on the screen, seen in Fig. 1 and Fig. 2) and use this pointer as input for model 2.
Model 2 has been trained with raw data and a variably sized pointer, the size of which correlates to the desired segmentation area. The user is only shown the final segmentation result from model 2.
Training and evaluation
For training and evaluation, we used an openly available dataset of abdominal CT scans and segmented lymph nodes from TCIA [7]. The test set was made of middle slices from 62 lymph nodes from 25 CT scans and 100 other CT scans were used for training. Many test slices had several lymph nodes visible, therefore, an image of a smoothed mask (1-second duration) was shown to highlight which lymph node to measure (Fig. 1 and Fig. 2).
DG segmentations and caliper-based measurements of long and short axis were performed by two radiologists (TS and HJ). This was done with a randomized order of lymph nodes to prevent recall bias.
Long and short axis measurements were automatically calculated from the DG-based segmentations. These measurements were compared to ground-truth masks in the downloaded dataset.