Using multiple open-access models trained on public datasets, we developed Tri-Reader, a comprehensive, freely available pipeline that integrates lung segmentation, nodule detection, and malignancy classification into a unified tri-stage workflow. The pipeline is designed to prioritize sensitivity while reducing the candidate burden for annotators. To ensure accuracy and generalizability across diverse practices, we evaluated Tri-Reader on multiple internal and external datasets as compared with expert annotations and dataset-provided reference standards

@misc{tushar2026trireaderopenaccessmultistageai,
      title={Tri-Reader: An Open-Access, Multi-Stage AI Pipeline for First-Pass Lung Nodule Annotation in Screening CT}, 
      author={Fakrul Islam Tushar and Joseph Y. Lo},
      year={2026},
      eprint={2601.19380},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2601.19380}, 
}

Tri-Reader tri-stage workflow for first-pass lung nodule annotation in screening CT. After lung segmentation (VISTA3D) removes extra-pulmonary candidates, Stage 1 performs consensus CADe using two complementary detection models (CADe-ROIonly and the false-positive–aware CADe-FPaware trained via strategic hard-negative mining); candidates detected by both are promoted to confidence = 1.0. Stage 2 applies ensemble malignancy scoring to disagreement candidates using two CADx classifiers trained on complementary distributions (DLCS24-CADxr50SWS and LUNA25- CADxr50); candidates with an averaged $CADx ≥ 0.10$ are promoted (confidence = 0.5). Stage 3 retains remaining candidates meeting $CADe ≥ 0.20$ (confidence = 0.2), yielding a single merged candidate list with tiered confidence intended to reduce annotator workload while preserving sensitivity; an optional rule-based NLP module can link report-derived descriptors to candidates when radiology reports are available.

Below candidate review panel for SAMPLE_0124 (68-year-old; Intgmultiomics cohort, SAMPLE_0124.nii). The leftmost tile is the dataset-provided ground-truth nodule; the remaining tiles are additional candidates proposed by our pipeline. Each tile shows the axial CT slice (left) and a magnified ROI (right); the yellow square marks the candidate location. The header above each tile reports ANod (annotation-consensus score), CADx (malignancy probability), CADe (detector confidence), Lrg.ax (largest axis, mm), lobar location, and slice index. Frame color follows the cancer-risk scale shown in the bottom bar (<0.25, ≥0.25, ≥0.50, ≥0.75).

• [1] 09/30/2025 - 📢 Created Github and Huggingface repo
• [2] 28/01/2025 - 📂 Mnauscript (Pre-print)
• [3] - 📂 Public release of 19K NLST CT Annotations .
• [4] - 📂 Public release of CT RATE Validation Dataset Annotations .