publications
No formally published papers yet. This page records ongoing work, the infrastructure supporting it, and the experiment methodology I follow.
ONGOING WORK
Interpretable pulmonary nodule classification and malignancy risk grading from chest CT
Provincial undergraduate innovation training project · approved 2026.05, ongoing · data processing and model experiments
The aim is classification and malignancy risk grading of pulmonary nodules on chest CT, with interpretable evidence for each decision. Public datasets such as LIDC-IDRI are the starting point. The current focus is the stability of the evaluation protocol: fixed data splits, seeds and evaluation scripts so that comparisons between configurations are meaningful, with method and conclusion kept separate.
PyTorch 3D medical imaging Interpretability
Breast ultrasound classification with cross-dataset external validation
2026.03 – 2026.07 · algorithms and software implementation
A benign/malignant classification baseline trained on BUS-BRA and evaluated on the independent BUSI test set, with the emphasis on cross-dataset generalisation rather than gains on a single dataset. Evaluation reports AUC, sensitivity, specificity, precision and F1 under a fixed threshold policy; the model exposes Grad-CAM attention maps as interpretability evidence.
External validation Grad-CAM Medical imaging
Fluorescence-guided osteomyelitis imaging software system
2026.06 – 2026.08 · independently implemented
An end-to-end chain from image processing to a clinical workstation interface for fluorescence-guided osteomyelitis imaging: registration and display of white-light and fluorescence views, lesion annotation and 3D review. The research interest lies in how algorithm outputs are organised into a clinically usable workflow, and in the engineering constraints of a restricted delivery environment (offline, physical media, heterogeneous hardware).
Three.js Medical software Delivery
RESEARCH INFRASTRUCTURE
Literature corpus knowledge base and ingestion pipeline
2026.08
A local knowledge base built to make literature review more efficient: content-hash idempotency, a task table driving resumable execution, and mixed local-GPU and cloud inference scheduling. 30,031 papers were processed, completing 30,018 (99.96%) in 11.2 hours — about 2.5 seconds per paper. Every knowledge entry carries evidence links back to the source paper, with a pending-review queue promoted by human checking, so retrieved results remain traceable.
Retrieval augmentation Pipeline engineering Evidence trail
Research workflow tooling and experiment management
2026.07 – present
Staged tooling for model development and paper writing: 13 skill modules and about 20 deterministic scripts with 90+ accompanying tests, covering topic selection, training, validation, writing and figure generation. On the experiment side, a single record template and an experiment registry keep configurations, data splits and conclusions together.
Workflow Experiment management Reproducibility
DATA AND SETUP
| Stage | Configuration |
|---|
| Public datasets | BUS-BRA and BUSI (breast ultrasound), LIDC-IDRI (chest CT nodules) |
| Framework | PyTorch |
| Local compute | RTX 5060 Laptop (8 GB VRAM) with 32 GB RAM, Linux under WSL2 |
| Server | Self-hosted Ubuntu machine for inference services and batch experiments |
| Experiment management | Fixed splits and seeds; a registry recording every configuration and conclusion |
The 8 GB VRAM budget requires trade-offs between batch size, mixed precision and model size, which keeps lightweight approaches a standing interest.
METHOD
- Decision criteria in advance: the threshold is set and recorded before the run, and never adjusted afterwards
- Failed runs recorded in the same format: for example, an attention routing module that dropped accuracy from 77.23% to 56.78% and collapsed mid-training is fully archived (see notes)
- Experiment registry: configurations, splits and conclusions logged for later comparison
- External validation first: metrics on the source dataset say little about generalisation
- Metrics interpreted per scenario: in medical imaging, sensitivity and specificity usually deserve more discussion than accuracy, and thresholds are reported alongside the numbers
MANUSCRIPT PLANS
- Turn the pulmonary nodule experiments into a reproducible record: data splits, evaluation protocol, ablations and failed runs preserved
- Once the evaluation protocol is stable, write up the method details and consider submission
For experiment details or code, see the projects page or email me directly.
