Deep Ensemble Models for Probabilistic CTV Breast Segmentation

A study evaluates six deep learning models for automatic clinical target volume segmentation in whole-breast radiotherapy using 961 planning CTs. The research leverages top-performing models to construct probability maps, aiming to improve consistency and mitigate bias in clinical predictions. Supported by the Italian Ministry of Health (CCM 2024), the dataset is a 2.9 MB PDF published by Cecilia Riani on figshare.

Use Cases

• Benchmarking deep learning segmentation models based on the evaluation of six architectures described in the study.
• Developing uncertainty quantification methods based on the model-based probability maps for clinical target volumes.
• Training or validating medical image segmentation algorithms based on the described dataset of 961 planning CTs for breast radiotherapy.

Strengths

• Model performance is comparable to inter-observer variability, with a reported Dice score of 0.90.
• The study is based on a cohort of 961 patient planning CT scans (861 for training, 100 for temporal testing).
• Includes a systematic evaluation of six advanced deep learning models and the creation of uncertainty probability maps.

Limitations

• The primary data file is a 2.9 MB PDF; the underlying image data (CT scans) and model outputs are not directly included.
• Data originates from a single institute, which may limit generalizability.
• Column-level documentation for any underlying tabular evaluation data is absent.

Provenance

Source

figshare, authored by Cecilia Riani.

Collection Method

Planning CTs from a single institute were used to train and evaluate deep learning models.

Freshness

Last updated 2026-04-30 05:24:07