Robust automatic soft tissue flap segmentation using a challenging case-enriched nnU-Net in head and neck CT images.
Authors
Affiliations (9)
Affiliations (9)
- LaTIM, INSERM, UMR 1101, Univ Brest, Brest, France.
- LITIS UR 4108, Henri Becquerel Cancer Center, Rouen, France.
- Siemens Healthineers, Courbevoie, France.
- Department of Radiation Oncology, Amiens, France.
- Department of Radiation Oncology, Institut Curie, Paris, France.
- GORTEC (Groupe d'oncologie radiothérapie tête et cou), Tours, France.
- LaTIM, INSERM, UMR 1101, Univ Brest, Brest, France. [email protected].
- Department of Radiotherapy, Comprehensive Cancer Center François Baclesse, Caen, France.
- ENSICAEN, CNRS/IN2P3, LPC Caen, UMR6534, Université de Caen Normandie, Caen, France.
Abstract
Reconstructive surgery with a flap makes the definition of postoperative radiotherapy volumes challenging. It may also result in errors in automatic segmentation atlases of organ-at-risk and nodal volumes. Automating flap segmentation process could assist clinicians in planning radiotherapy and enable characterization of flap evolution over time and after radiotherapy. Flaps vary significantly in shape, volumes and associated artefacts. We therefore enriched a previously built training dataset with challenging cases to obtain a more robust real-world flap segmentation. Within the framework of the state-of-the-art nnU-Net deep learning architecture, we investigated whether constructing a training dataset with enhanced representation of challenging cases, often associated with poor segmentation performance or outright failures, could improve the overall accuracy and robustness of automated flap segmentation, based on Dice scores compared through paired Wilcoxon signed-rank tests. Clinical trial and real-world data were selected to increase the heterogeneity and enrich the training set with rare challenging cases (such as pedicled flaps, small flaps, unusual location including maxillary flaps, bone resection, presence of dental artefacts or bite block). This enriched training dataset led to improved performance of the nnU-Net model, increasing the mean Dice scores from 0.66 ± 0.29 to 0.74 ± 0.20 (p < 0.001), with median Dice scores rising from 0.76 to 0.80. Robust flap segmentation was achieved without modifying the neural network architecture, loss function, or algorithmic structure, through enrichment of the training set with anatomically and visually challenging cases. This model can be used for detailed analysis of geometrical and textural flap changes over time.