DeepNerve-3D
这个项目的价值不在炫界面，而在把困难的医学影像问题讲清楚：3D CBCT 中下牙槽神经前景极少、对比度低、拓扑连续性要求高。DeepNerve-3D 用 SwinUNETR、滑窗推理和连通域约束，把研究思路落到可运行的分割 pipeline 里。
Mandibular nerve segmentation in CBCT is difficult because the target is tiny, low-contrast, and topologically continuous; a visually plausible mask can still fail if the nerve path breaks.
I used a SwinUNETR-based 3D segmentation pipeline with sliding-window inference, class-imbalance awareness, and topology-oriented post-processing constraints.
The project adds a credible medical-imaging research case to the portfolio and shows that the AI direction is broader than web products alone.
- 01Built the research-style training and inference pipeline.
- 02Framed the task around foreground scarcity, topology continuity, and compute limits.
- 03Presented the work as an engineering prototype with explicit constraints rather than a clinical claim.
Medical AI prototypes should be judged by problem framing, constraints, and reproducible pipelines. The goal is not to overclaim clinical performance, but to show disciplined research engineering.
PyTorch and MONAI pipeline for 3D CBCT volumes. SwinUNETR captures long-range anatomical context, sliding-window inference handles memory limits, and connected-component logic reduces fragmented false positives in post-processing.
- 01Topology matters because a locally plausible mask can still be wrong if the nerve path breaks.
- 02Sliding-window inference and post-processing make the pipeline practical under limited GPU memory.
- 03The case adds research depth to the applied product projects without overclaiming clinical readiness.