Comparing Deep Learning Models for Cancer Segmentation
Author Information
Author(s): Chen Xinru, Zhao Yao, Baroudi Hana, El Basha Mohammad D., Daniel Aji, Gay Skylar S., Yu Cenji, Wang He, Phan Jack, Choi Seungtaek L., Goodman Chelain R., Zhang Xiaodong, Niedzielski Joshua S., Shete Sanjay S., Court Laurence E., Liao Zhongxing, Löfman Fredrik, Balter Peter A., Yang Jinzhong
Primary Institution: The University of Texas MD Anderson Cancer Center
Hypothesis
How do custom-trained deep learning models compare to vendor-pretrained models in segmenting cancers?
Conclusion
Custom-trained models significantly improve segmentation accuracy for head-and-neck, breast, and prostate cancers compared to vendor-pretrained models.
Supporting Evidence
- Custom-trained models showed significant improvements over vendor-pretrained models for 14 of 24 organs at risk.
- The average Dice similarity coefficients for custom-trained models were higher than those for vendor-pretrained models.
- Statistically significant differences were found in segmentation accuracy across various cancer types.
Takeaway
This study shows that using a small amount of local data can help create better computer models for identifying cancer areas in scans.
Methodology
The study used clinical CT scans and contours from 210 patients to train and validate segmentation models, comparing vendor-pretrained and custom-trained models.
Potential Biases
Potential variability in manually delineated contours could affect model accuracy.
Limitations
The study did not include subjective evaluations by physicians and did not assess the optimal number of patients required for custom training.
Participant Demographics
210 patients: 53 with head-and-neck cancer, 49 with left breast cancer, 55 with right breast cancer, and 53 with prostate cancer.
Statistical Information
P-Value
p<0.001
Statistical Significance
p<0.001
Digital Object Identifier (DOI)
Want to read the original?
Access the complete publication on the publisher's website