OSCC PNI patient-level MIL checkpoints
Inference-only PyTorch state dictionaries for the OSCC PNI research demonstration. The collection includes Swin
Transformer MIL, Vision Transformer MIL, DenseNet-121 MIL, and the currently supplied ResNet checkpoint. Training
optimizer states were removed without quantization. manifest.json records SHA-256 checksums of the original training
checkpoints and these inference exports.
Interactive demo
The public ZeroGPU Gradio application is available at OSCC PNI MIL Research Demo. It is for research and reproducibility only; upload only de-identified prepared images that you are authorized to process.
Source code
Deployment-ready Gradio source is available in the
space/ directory. It includes the app,
dependency lock, model definitions, Swin implementation, and inference-checkpoint export utility. The app uses local
files under weights/ when present and otherwise downloads the matching checkpoint from this repository.
Temporary ResNet checkpoint notice
The public interface is labelled ResNet-101 MIL to match the manuscript terminology, but resnet152_mil.pt
currently contains the supplied ResNet-152 checkpoint. It is a deployment placeholder and must be replaced with the
study ResNet-101 checkpoint before the ResNet result is cited or the repository is treated as the final archival
implementation.
Intended use
Research and reproducibility only. These models are not medical devices and are not intended for diagnosis, triage, prognosis, or treatment decisions. The underlying retrospective study has not completed prospective site-separated external validation.
Inputs
The associated prototype expects de-identified, ROI-guided tumor-bearing contrast-enhanced CT axial slices prepared as PNG/JPEG, resized to 224 × 224, converted to RGB, and normalized with ImageNet mean and standard deviation. Raw DICOM series and unprepared whole-volume CT are not supported by the demonstration.