ProstatID™ is a software-only Computer-Aided Detection (CADe), and Diagnosis (CADx) system for use by a trained physician (typically a radiologist) to aid in the detection and diagnosis of prostate cancer using post-processed Magnetic Resonance (MR) images of the prostate.

The system is not installed at the user’s site, but rather acts as a software-as-a-service (SaaS) where the user sends prostate studies to Bot Image, the studies are automatically processed, and the results are sent back to the user.

History of the

“Detection of prostate cancer in multiparametric MRI using random forest with instance weighting,” Journal of Medical Imaging, 2017.

  • National Institutes of Health (NIH)
  • Author: Nathan Lay
  • Co-authors: Dr. Baris Turkbey, Dr. Peter Choyke, Dr. Ronald Summers, et. al.

Trained Random Forest model to detect highly suspicious lesions in the prostate using T2w, DWI, and ADC image sequences and 64 image features

Original NIH Algorithm Limitations:

  1. Patient Data → 224 patient cases
    • single institution
    • single MR scanner (Philips Achieva 3.0T)
    • trained algorithm may not translate to other scanners or protocol
  2. Relied on a commercial software for segmentation of the prostate anatomy, which involved manual correction.
  3. Model was trained to detect highly suspicious lesions, regardless if the biopsied lesion came back negative.
    • trained to act like the radiologist, not to perform better than the radiologist
    • potentially increasing false positives with us
  4. Method was developed in the research setting
    • high regard for good results, less regard for good software design practices
    • ad hoc system not designed for marke

ProstatID: from
Research to Market

  • ScanMed, LLC acquired the rights to the NIH data and source code.
  • Launched software company, Bot Image, dedicated to developing algorithms for the medical field.
  • Improved NIH algorithm
    • a. ScanMed, LLC acquired the rights to the NIH data and source code.
      • i. multiple sites, multiple MR scanners
    • b. Developed automated segmentation method for prostate anatomy.
      • i. utilizes 3-D convolutional neural networks
    • c. Random forest model trained on true outcomes.
      • i. benign biopsy results considered as true negatives, not positives like previous NIH model
    • d. Additional development for incorporation into the clinical setting.
      • image quality checks
      • input/output compliant with the Digital Imaging and Communications in Medicine (DICOM) format
      • security and high availability
  • Clinical Setting: post-processing MRI prostate studies
  • Processing Pipeline

FDA Regulatory Approval Pathway

    • Predicate Device: TransparaM by Screen Point Medical B.V.
      • FDA-cleared software product for diagnosis of breast cancer lesions
      • same intended use as ProstatID:
        • “Radiological computer-assisted detectionand diagnostic (CAD) software for lesions suspicious for cancer”
        • allows for a 510(k) approval process
    • FDA 510(k) Submission
      • Standalone Performance Assessment
        • The developed ProstatID software was tested internally to evaluate the correlation of its outputs to a reference standard.
        • Tested on 150 patient cases set aside prior to algorithm training.
      • Clinical Performance Assessment
        • Evaluated the effect of using the ProstatID software in the clinical setting.
        • Tested on a separate set of 150 patient cases selected prior to training.
        • Recruited 12 physicians to participate:
          • 1st Read: current standard of care without ProstatID
          • Tested on a separate set of 150 patient cases selected prior to training.
          • 2nd Read: current standard of care with ProstatID
    • Standalone Performance Assessment
      • ROC analysis of ProstatID index value and true outcomes
      • 3mm regions of entire prostate region analyzed
      • 150 patient cases → 77 benign, 73 cancer
      • Results: overall AUC = 0.852 [0.816, 0.888]
        • also broken down into different MR field strength
    • Standalone Performance Assessment
      • Multiple reader, multiple case (MRMC) study design
      • 12 physicians reading 150 cases to achieve statistical power of 0.80
      • 150 retrospective patient cases → 80 benign, 70 cancer
      • Readers instructed to give each detected lesion a score 1-5 (PI-RADS v2)
      • Results: average change in AUC = +0.061 [0.016, 0.105] (p = 0.0097)
        when using ProstatID
        • Improved detection accuracy by 27.8% using FROC analyses
        • Decreased False Positive (biopsies) by 36%

Next Steps

Continuously Improve Detection Accuracy
Random forest has its limitations, so train more sophisticated models (Deep Learning):
neural network using the same 64 image features as the random forest model
RF model: AUC = 0.662 on 191 biopsied suspicious lesions
NN model: AUC = 0.739 on the same lesions → 10% increase in accuracy
convolutional neural network using images themselves as input
Improve Reference Standard (Ground Truth)
Targeted biopsy point locations → Confirmation with whole-mount slides after prostatectomy
UCLA algorithm (FocalNet)
Migration to Non-Invasive Cancer Grading with MRI
Predicting Gleason Score by utilizing histopathology slides and training algorithm
Currently working on a small business innovation research (SBIR) applications