Maximilian Dressler, M.D.

Can published station-level prevalence be turned into a transparent surgical decision-support prototype?

Interactive descriptive prototype for 11 perigastric stations using systematic-review T-stage prevalence as the professor-facing evidence layer; individualized calibration is future work.

Literature-informed scaffold; calibration pending

⚠ All values are placeholders — not validated for clinical use. Real validation requires SNUH/KLASS registry data.

View code →

Can a surgical-video segmentation pipeline move from laparoscopic benchmark data toward robotic gastrectomy instrumentation?

Validated leave-videos-out CholecSeg8k benchmark work, a public gastrectomy domain-shift demo, and a preliminary SISVSE transfer-learning run using public robotic distal-gastrectomy frames.

CholecSeg8k CV: mean IoU 0.7574 · Dice 0.8614 (DeepLabV3; U-Net vs DeepLabV3 difference not statistically significant, p≈0.21) | SISVSE Run B: zero-shot IoU 0.0642 → fine-tuned IoU 0.4913

⚠ Preliminary public-dataset evidence only — not a validated clinical model, not trained on private SNUH data, and not suitable for clinical use.

View code →

0.7574 CholecSeg8k mean IoU

0.8614 CholecSeg8k mean Dice

Evidence	Dataset / source	What it shows
5-fold leave-videos-out CV	CholecSeg8k · 8,080 annotated frames / 17 videos	Reproducible benchmark proof-of-method for laparoscopic instrument segmentation
Public gastrectomy clip demo	PMC3752361 · 150 sampled frames	Qualitative domain-shift check on public laparoscopic gastrectomy footage
SISVSE Run B transfer	Public SISVSE robotic distal-gastrectomy frames · N=480 test frames	Zero-shot transfer was poor; targeted SISVSE fine-tuning substantially improved segmentation

SISVSE setup	Macro IoU	Macro Dice	N	Interpretation
Zero-shot CholecSeg8k → SISVSE	0.0642	0.1148	480	Strong laparoscopic→robotic domain gap
Fine-tuned on SISVSE	0.4913	0.6260	480	Preliminary transfer signal after target-domain adaptation

Fine-tuned SISVSE Run B overlays

Representative public-dataset examples: one success, one typical case, and one failure case. Included deliberately to show both capability and limits.

The SISVSE result is intentionally framed as preliminary public-dataset transfer evidence: zero-shot transfer from CholecSeg8k to robotic gastrectomy frames failed badly (macro IoU 0.0642), while target-domain fine-tuning improved macro IoU to 0.4913 and macro Dice to 0.6260 on 480 held-out frames. This supports the feasibility of the technical workflow, but clinical validation would still require permissioned SNUH robotic gastrectomy data, IRB/privacy review, and prospective evaluation.

How do we ensure AI predictions are clinically trustworthy, not just accurate?

Implementation of validation methodology principles inspired by TRIPOD/PROBAST: discrimination (C-index), calibration (Brier score, ECE), and survival analysis linkage (Cox regression, Kaplan-Meier). Decision-curve analysis is planned but not yet implemented.

Survival analysis pipeline tested on TCGA STAD; prediction and calibration models demonstrated on synthetic data

⚠ Demonstrates validation methodology, not novel algorithms. Clinical validation requires prospective SNUH data.

View code →

What is the survival landscape across gastric cancer stages?

Reproducible pipeline generating TNM heatmaps, Kaplan-Meier curves, and stage distributions from TCGA STAD PanCanAtlas data, with Benjamini-Hochberg correction for multiple comparisons.

6-fold survival differential documented (Stage IA: 72.2 mo vs Stage IV: 12.0 mo)

⚠ Based on public TCGA data — KLASS registry data needed for clinical applicability.

View code →