The group

Five companies, one engineering standard.

Each company runs its own practice and answers to its own market. What they share is an engineering standard and an infrastructure spine.

All companies

Arioron Developers Software engineering practice Arioron Holdings Group ownership and capital Arioron Labs Applied research Vortarz Network Game development and server hosting Elakar Khobor Local news platform

Arioron Developers Arioron Holdings Arioron Labs Vortarz Network Elakar Khobor All companies

Appearance

← Research

ARL-2025-009 · Medical Imaging

OncoDetect Titan: robustness over benchmark fit in lung cancer detection

A multi-model ensemble for lung cancer detection, trained explicitly on degraded and inconsistent scan data. We argue that robustness to real hospital conditions matters more than benchmark position, and report sensitivity on an external test set.

Results ARL-2025-009

Sensitivity, external set Recall on held-out data

100.0%

Ensemble members Independently trained networks

6.0

In short

Trained on deliberately degraded scans, not curated cohorts
Six-network ensemble rather than a single model
Sensitivity alone is not clinical validation
Decision support under supervision, not a diagnostic product

Diagnostic models tend to be evaluated under conditions no hospital actually operates in: consistent scanners, clean acquisition, curated cohorts.

OncoDetect Titan was built the other way round. We trained an ensemble of six networks on deliberately degraded data — older scanners, inconsistent slice thickness, noise — on the assumption that a model which only works on clean inputs is not a clinical tool.

Evaluation

On an external test set the ensemble reaches 100% sensitivity. That number requires immediate context: sensitivity alone is not a clinical validation, the external set is finite, and specificity and prospective evaluation are the work that follows.

Intended use

This is decision support, not diagnosis. The relevant deployment question is whether it helps a radiologist in a hospital with older equipment and a heavy caseload, and that question is answered in clinical settings rather than on a benchmark.

Related research

ARL-2025-011 Vex-Amber-Fable-2.0: parameter efficiency in the sub-3B class A 2-billion-parameter causal language model trained for software engineering and reasoning tasks, evaluated against models an order of magnitude... Aug 2026 ARL-2025-004 Vex-Amber-Mini: reasoning at 0.6B parameters A 0.6-billion-parameter model optimised for code generation and general text tasks on constrained hardware, with notes on the distillation... Apr 2026