Related Experiment Video
Updated: Jun 21, 2026
07:13
Comparison of Predictive Performance of Three Lymph Node Staging Systems in Colorectal Signet Ring Cell Carcinoma Based on Machine Learning Model
Published on: April 18, 2025
Benchmarking deep learning models for laryngeal cancer staging using the LaryngealCT dataset
Nivea Roy1,2, Son N Tran1, Atul Sajjanhar1
1School of Information Technology, Deakin University, Burwood, VIC, Australia.
Scientific Reports
|June 19, 2026
Summary
A new dataset, LaryngealCT, provides standardized computed tomography (CT) scans for developing reproducible deep learning (DL) models in laryngeal cancer research. This benchmark enables AI advancements for better clinical decisions.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Oncology
Background:
- Laryngeal cancer research requires standardized datasets for reproducible deep learning (DL) model development.
- Existing imaging data is fragmented, hindering AI-driven advancements in diagnosis and treatment planning.
Purpose of the Study:
- To introduce LaryngealCT, a comprehensive, curated benchmark dataset of 1,029 laryngeal CT scans.
- To evaluate the performance of six 3D deep learning architectures for laryngeal cancer classification tasks.
- To provide open-source data and tools to foster reproducible AI research in laryngeal oncology.
Main Methods:
- Aggregated 1,029 CT scans from six TCIA collections into the LaryngealCT benchmark.
- Extracted uniform 1mm isotropic larynx volumes using a validated, weakly supervised parameter search.
- Benchmarked six 3D DL models (3D CNN, ResNet variants, DenseNet121) on early vs. advanced and T4 vs. non-T4 classification tasks.
- Assessed model explainability using GradCAM++ with thyroid cartilage overlays.
Main Results:
- The custom 3D CNN achieved the highest performance (Accuracy=0.854, F1-macro=0.841) for early vs. advanced laryngeal cancer classification.
- For T4 classification, most models exceeded 0.82 AU-ROC, but T4 sensitivity was limited (≤0.412), with ResNet101 showing improved recall (0.706).
- Explainability analysis revealed plausible, albeit modest, peri-cartilage activations for T4 classification.
Conclusions:
- LaryngealCT provides a standardized, open-source resource for reproducible AI research in laryngeal cancer.
- The benchmark facilitates the development and validation of deep learning models for improved laryngeal cancer detection and staging.
- This initiative supports AI-driven clinical decision-making in laryngeal oncology through accessible data and tools.
