Learning from a Few Shots: Data-efficient Cervical Vertebral Maturation Assessment

Helen Schneider, Aditya Parikh, Priya Priya, Maximilian Broß, Tom Verhofstadt, Anna Konermann, Rafet Sifa

Keywords: Few-shot Learning, Transfer Learning, Orthodontic, CVM Assessment

TL;DR: Data-efficient Deep Learning for CVM Assessment

Abstract:

The timing of treatment is a crucial decision in orthodontics. Initiating treatment during
the appropriate growth phase leads to optimal patient outcomes and can prevent prolonged
treatment durations. The most commonly used method for classifying growth phases is
cervical vertebral maturation (CVM) assessment, which categorizes CVM into six stages
based on the shape and size of the cervical vertebrae. Due to the complexity of manual CVM
analysis, it often falls short in performance when assessed visually. Deep learning methods
can assist physicians in classifying CVM stages, thus improving orthodontic workflows and
treatments. However, a significant challenge in deep learning-based CVM assessment is
the limited dataset volume, resulting from difficulties in data collection and annotation.
While small training datasets can greatly hinder the model’s generalization performance,
research on data-efficient training methods for CVM assessment is still lacking. To the best
of our knowledge, this paper is the first to evaluate the potential of few-shot learning and in-
domain transfer learning for CVM assessment. Specifically, we investigate the architectures
ResNet18 and MedSam-2D. Few-shot learning enhances classification performance by up
to 9%. Additionally, in-domain pre-training (using chest X-ray data) results in a significant
performance increase of up to 4%.

Primary Subject Area: Learning with Noisy Labels and Limited Data

Secondary Subject Area: Application: Other

Paper Type: Validation or Application

Registration Requirement: Yes

Midl Latex Submission Checklist: Ensure no LaTeX errors during compilation., Created a single midl25_NNN.zip file with midl25_NNN.tex, midl25_NNN.bib, all necessary figures and files., Includes \documentclass{midl}, \jmlryear{2025}, \jmlrworkshop, \jmlrvolume, \editors, and correct \bibliography command., Did not override options of the hyperref package, Did not use the times package., All authors and co-authors are correctly listed with proper spelling and avoid Unicode characters., Author and institution details are de-anonymized where needed. All author names, affiliations, and paper title are correctly spelled and capitalized in the biography section., References must use the .bib file. Did not override the bibliographystyle defined in midl.cls. Did not use \begin{thebibliography} directly to insert references., Tables and figures do not overflow margins; avoid using \scalebox; used \resizebox when needed., Included all necessary figures and removed *unused* files in the zip archive., Removed special formatting, visual annotations, and highlights used during rebuttal., All special characters in the paper and .bib file use LaTeX commands (e.g., \'e for é)., Appendices and supplementary material are included in the same PDF after references., Main paper does not exceed 9 pages; acknowledgements, references, and appendix start on page 10 or later.

Submission Number: 171