Compass: Prostate Cancer Detection Needs Multi-View Context

Compass is a multi-instance learning model for case-level prostate cancer detection from micro-ultrasound sweeps and systematic biopsy cores.

Installation

git clone https://github.com/mharmanani/Compass
cd Compass
pip install -e .

Requirements

• Python >= 3.9
• PyTorch >= 2.0
• See setup.py for full dependency list

Environment Variables

Variable	Required	Description
PNF_PLUS_CHECKPOINT	Yes	Path to the ProstNFound+ pretrained backbone checkpoint (MedSAM-based)
WANDB_API_KEY	No	WandB API key for experiment logging

The ProstNFound+ checkpoint provides the image encoder backbone. It is a MedSAM checkpoint fine-tuned on micro-ultrasound data and is available upon request from the authors.

Data

Compass was trained on the OPTIMUM micro-ultrasound dataset, which is a proprietary clinical dataset and is not included in this repository.

The dataset is specified via a samples.json file. The top level is a JSON object keyed by case ID. Each case has the following structure:

{
  "UA-004": {
    "case": "UA-004",
    "center": "UA",
    "age": 67,
    "psa": 6.5,
    "max_grade_group": 2.0,
    "study_arm": "MicroUS",
    "cines": [
      {
        "cine_id": "UA-004-001",
        "image_path": ["/path/to/core_001.png"],
        "roll_path": "/path/to/core_001_roll.npz",
        "needle_mask_path": "/path/to/core_001_mask.png",
        "GG": 2.0,
        "% Cancer": 40.0,
        "PRI-MUS": "4"
      }
    ],
    "sweeps": [
      {
        "sweep_id": "UA-004-S001",
        "sweep_path": "/path/to/sweep_001.h5",
        "roll_path": "/path/to/sweep_001_roll.npz"
      }
    ]
  }
}

Case fields:

• case (str): case identifier (matches the top-level key)
• center (str): imaging site
• age (int): patient age at biopsy
• psa (float): PSA value (ng/mL)
• max_grade_group (float): case-level Gleason grade group (0 = benign)
• study_arm (str): study protocol arm

Biopsy core fields (cines list):

• cine_id (str): unique core identifier
• image_path (list of str): path(s) to biopsy B-mode PNG; the last path is used
• roll_path (str): path to .npz file containing a roll key — a 1-D array of probe roll angles (degrees) captured during the biopsy; the last frame is the canonical angle
• needle_mask_path (str, optional): path to PNG needle segmentation mask
• GG (float): Gleason grade group (NaN or 0 = benign)
• % Cancer (float): percentage of core involved by cancer (NaN = benign)
• PRI-MUS (str/int, optional): PRIMUS micro-ultrasound suspicion score (1–5)

Sweep fields (sweeps list):

• sweep_id (str): unique sweep identifier
• sweep_path (str): path to HDF5 file; must contain a sweep dataset of shape (T, H, W) or (T, H, W, C) — a sequence of B-mode frames
• roll_path (str): path to .npz file containing a roll key — a 1-D array of probe roll angles (degrees) over the sweep

The case-level cancer label (label) is derived automatically: a case is positive (label = 1) if any biopsy core has GG ≥ 2 (clinically significant cancer).

The patient splits file (resources/splits.json) defines cross-validation folds using case IDs.

Training

python train.py --config configs/compass.yaml \
    -o runtime.fold=0 -o runtime.group_id=run1

To run all 5 folds:

for fold in 0 1 2 3 4; do
    python train.py --config configs/compass.yaml \
        -o runtime.fold=$fold -o runtime.group_id=run1
done

Configuration options can be overridden with -o key=value.

Model Architecture

Compass uses:

• ProstNFound+ as a frozen image encoder backbone (MedSAM-based)
• A BERT-based MIL aggregator to integrate biopsy core features
• A sweep video encoder for multi-view temporal context
• Clinical prompt conditioning (PSA, age)
• Sinusoidal roll positional embeddings

Citation

@inproceedings{wilson2026compass,
  title   = {Compass: Prostate Cancer Detection Needs Multi-View Context},
  author  = {Wilson, Paul F. R. and Harmanani, Mohamed and Guo, Zhuoxin and Dzikunu, Obed K. and Cash, Hannes and Kinnaird, Adam and Wodlinger, Brian and Abolmaesumi, Purang and Mousavi, Parvin},
  booktitle = {International Conference on Medical Image Computing and Computer-Assisted Intervention (MICCAI)},
  year    = {2026},
}

License

MIT License

Copyright (c) 2026 Paul F. R. Wilson, Mohamed Harmanani, Zhuoxin Guo, Obed K. Dzikunu, Hannes Cash, Adam Kinnaird, Brian Wodlinger, Purang Abolmaesumi, and Parvin Mousavi

Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions:

The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.

THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.