Task-Specific Fine-Tuning for Interactive Deep Learning Segmentation for Lung Fibrosis On CT Post Radiotherapy
MJ Trimpl1,2,3*, P Salome4,5,6,7, D Walz4,5,6,7, J Hoerner-rieber6,7, S Regnery6,7, EPJ Stride1, KA Vallis3, J Debus6,7, A Abdollahi4,5,6,7, MJ Gooding2, M Knoll4,5,6,7, (1) Institute of Biomedical Engineering, Department of Engineering Science, Oxford, GB, (2) Mirada Medical, Oxford, GB, (3) Oxford Institute for Radiation Oncology, Oxford, GB, (4) CCU Translational Radiation Oncology, German Cancer Consortium (DKTK) Core-Center Heidelberg, Heidelberg University Hospital (UKHD) and German Cancer Research Center (DKFZ), Heidelberg, DE, (5)Division of Molecular and Translational Radiation Oncology, Heidelberg Faculty of Medicine (MFHD), Heidelberg University Hospital (UKHD), Heidelberg Ion-Beam Therapy Center (HIT), Heidelberg, DE, (6) Department of Radiation Oncology, Heidelberg University Hospital, Heidelberg, DE, (7) National Center for Radiation Oncology (NCRO), Heidelberg Institute for Radiation Oncology (HIRO), Heidelberg, DE
Presentations
SU-E-BRB-8 (Sunday, 7/10/2022) 1:00 PM - 2:00 PM [Eastern Time (GMT-4)]
Ballroom B
Purpose: Lung fibrosis segmentation is needed for translational research to optimize treatment, but currently few segmentation tools exist. This study investigates the impact of fine-tuning a general interactive deep learning method on lung fibrosis segmentation given a small training set.
Methods: 211 Non-Small Cell Lung Cancer patients (T1-4N0-3M0-1) with Stereotactic Body Radiation Therapy were evaluated. In n=38 (18%), radiation-induced fibrosis was described in follow-up CT imaging and manually segmented to obtain a ground truth. A deep learning tool was evaluated against manually segmented fibrotic volumes using either original pre-trained weights (trained on larger datasets including 19 different organs/structures, excluding lung fibrosis) or after fine-tuning on the lung fibrosis cases using 5-fold cross-validation. To segment, the model receives one user interaction as an input, in the form of a single contour at the central axial slice of the structure. Additionally, the model is evaluated given multiple contour inputs at a fixed slice distance. Fine-tuning was performed using extensive image augmentation and pixel-wise binary cross-entropy with the Adam optimizer at an initial learning rate of 1e-6.
Results: Fine-tuning resulted in a significant increase in segmentation performance (p-value<0.01), when given a single input; increasing Dice Similarity Coefficient from 0.60±0.15 to 0.72±0.11, and reducing 95% Hausdorff Distance from 23±20mm to 15±8mm and relative Added Path Length from 0.63±0.11 to 0.51±0.12. The original model can achieve similar performance as after fine-tuning when given contour inputs every seven slices instead of a single contour.
Conclusion: The performance of the fine-tuned model with a single interactive input is comparable to the original model’s performance with ~4 inputs. This demonstrates the potential time savings provided by task-specific fine-tuning and continual learning for structures that still require manual segmentation.
Funding Support, Disclosures, and Conflict of Interest: This project has received funding from the European Union's Horizon 2020 research and innovation program under the Marie Sklodowska-Curie grant agreement No.766276. KAV acknowledges funding support from CRUK Radnet Centre (A28736).
Keywords
Segmentation, CT, Computer Vision
Taxonomy
IM/TH- Image Segmentation Techniques: Machine Learning
Contact Email