Purpose
Semi-automatic area and volumetric measurements of lymph nodes and tumors are known to be more reliable than caliper-based measurements [1-3]. They are however still laborious and not routine in clinical practice.
Deep learning-based tools using convolutional neural networks (CNNs) are considered state-of-the-art for automated image segmentation, but consistently accurate automated segmentation remains challenging due to the high variability of anatomy and pathology [4]. Manual or semi-automated tools that do not make use of CNNs remain standard for most segmentation tasks.
DeepGrow (DG) is an interactive...
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...
Results
We did not find any significant difference in the accuracy of long and short axis measurements between using traditional caliper-based measurements and DG-based measurements (see Fig. 3). Both raters had on average smaller errors using DG on both long and short axis measurements, but the differences were not statistically significant.
One rater (HJ) used more time per lymph node using DG than caliper-based measuring (10.5 s.vs. 9.0 s.), but the difference was not significant. The second rater used significantly less time with DG than caliper-based...
Conclusion
We have shown that interactive 2D segmentation with deep learning can be used for measuring lymph nodes with similar or better accuracy and speed compared to traditional caliper-based measuring.
Further, segmentation provides area measurements that may have additional value compared to long and short axis measurements, however, this potential additional value was not assessed in this study.
Our approach has been made available asan open web-based application [6]. We hope that our demonstrated application will facilitate developments of other deep learning-based implementations that use input...
Personal information and conflict of interest
T. Sakinis; Oslo/NO - CEO at Artificial Intelligence AS H. B. Jenssen; Oslo/NO - Advisory Board at Artificial Intelligence AS
References
1. Hoink AJ, Wessling J, Koch R, Schulke C, Kohlhase N, Wassenaar L, et al. Comparison of manual and semi-automatic measuring techniques in MSCT scans of patients with lymphoma: a multicentre study. European radiology. 2014;24(11):2709-18.
2. Fabel M, Bolte H, von Tengg-Kobligk H, Bornemann L, Dicken V, Delorme S, et al. Semi-automated volumetric analysis of lymph node metastases during follow-up--initial results. European radiology. 2011;21(4):683-92.
3. Arnljots TS, Al-Sharbaty Z, Lardner E, All-Eriksson C, Seregard S, Stalhammar G. Tumour thickness, diameter, area or volume? The prognostic...