Prostate Commander · Research Project · Diagnostic Imaging

AI-Based Analysis of Prostate MRI Sequences and Automated Lesion Detection

Prostate Commander is not just another generic DICOM viewer, but a lightning-fast, highly specialized analysis tool built for exactly one purpose: automatically contouring radiological cross-sectional images, presenting them to the user for review, and thereby speeding up the preparation of fusion biopsies and follow-up assessments – entirely on your own computer, with no cloud connection.

Not a DICOM Viewer – an Analysis Tool Lightning Fast Offline & Local Effortless: Drag & Drop Fully Automatic Sequence Detection Vendor-Independent PI-RADS-Oriented Lesion Search Manually Correctable Test Version Available

View Concept Get in Touch

Part 1 · Concept

What, Why, and For Whom

The software supports the evaluation of T2-weighted MRI scans of the prostate. It automates the most time-consuming part of preparation – tracing the organ boundaries slice by slice – and provides the result as a reviewable, correctable suggestion.

💡 The Motivation

This software grew out of daily clinical practice: as a radiologist who performs MRI/ultrasound fusion prostate biopsies together with a urologist, I was confronted with a wide variety of examinations from many different sources – of correspondingly varying quality, both in the scans themselves and in the accompanying reports. To speed up the upfront validation process and quickly gain an overview of how to plan the fusion biopsy most efficiently and safely, I wrote this program.

🎯 Purpose

Automated contouring of the prostate and, optionally, suspicious areas („lesions“) in axial T2 slices, as a basis for planning MRI/ultrasound fusion biopsies and for volume estimation.

⚙️ What It Does

Reads a DICOM series and fully automatically selects the required sequences (T2, diffusion). ADC maps are calculated independently from the diffusion images, as is the volume of the prostate and – if present – of a lesion. After AI-based contouring of the prostate, the software searches for lesions in the T2, diffusion, and ADC images – oriented on PI-RADS criteria. The complete diagnostic findings of the examination are summarized in a structured Word document.

👥 For Whom

For medical professionals in radiology and urology who assess prostate MRI series or prepare fusion biopsies and want to retain control over every automatically generated contour.

Core Principle: A Suggestion, Not an Automatism

Every automatically generated contour remains a suggestion. It is color-coded as active or inactive, and can be adjusted, deleted, or redrawn slice by slice. Slices on which the AI finds no plausible segment are automatically marked inactive, rather than forcing an incorrect contour.

Human retains final control
No automatic diagnosis
Traceable intermediate steps
Anonymous, local processing

Schematic illustration – not a real patient scan

Part 2 · Technical Background

How It's Built, Where It Runs

The application is designed as a standalone Windows program: no cloud processing, no transmission of image data to third parties. All computations run locally on the machine where it is started.

💻 Runtime Environment

Packaged as a standalone Windows application (no separate Python installation required). Runs fully offline on standard PC hardware. An available graphics card (CUDA) is automatically used for AI computation; otherwise the application falls back to CPU-optimized processing.

🧠 AI Component

A neural network specialized for medical cross-sectional images (deep-learning segmentation model) generates the suggested contour for each slice. The model can be fine-tuned as needed with additional, locally available training data.

📄 Image Format

Processes DICOM – the international standard for medical imaging data. Automatically detects the relevant axial T2 series within an examination folder and correctly assigns additional sequences (e.g. diffusion).

🔒 Data Handling

Image data never leaves the computer. Depending on the edition, intermediate results are either stored locally for later sessions or – in the lightweight edition – completely deleted after every session.

🛠️ Origin

Programmed through direct interaction between a human and the AI coding assistant „Claude Code“ by Anthropic – making the application itself also a product of artificial intelligence. Written in Python, runs both with and without CUDA GPU support. The logic behind the lesion search was likewise developed together with Claude Code based on the PI-RADS protocol and implemented in program code.

🧪 Test Version

A functionally limited, open edition is available for testing: loading exclusively via drag&drop, with no persistent data storage – all examination data is automatically and completely deleted after every session.

📋 Protocol Scope

Contrast-enhanced sequences were deliberately omitted: abbreviated MRI protocols without contrast agent are already imminent, so the lesion search is specifically designed around T2, diffusion, and ADC.

Simplified data flow within the application

Part 3 · Program Operation & Logic

From Import to Verified Contour

The workflow is deliberately kept linear: load, review, correct, evaluate. Each step is individually traceable and can be interrupted at any time.

Load Examination – Effortlessly A DICOM directory – whether from a CD, removable media, or folder – is simply dragged and dropped onto the application. No manual selection of files or sequences is required. CDs brought in by patients, or examinations made available for download by radiology institutes, are no problem either: Prostate Commander detects and imports – completely independently and regardless of the MRI scanner manufacturer – all sequences required for the analysis.
Automatic Sequence Detection & ADC Calculation The application automatically detects the required sequences (T2, diffusion). ADC maps are calculated independently from the diffusion images – with no user input required.
AI Contouring of the Prostate The AI component computes a suggested contour for the prostate on each slice. Slices with no plausible finding are automatically marked inactive.
PI-RADS-Oriented Lesion Search Within the segmented prostate contour, the application jointly searches the T2, diffusion, and ADC images for lesions – oriented on PI-RADS criteria.
Review & Correction The user reviews each slice on the axial image, adjusting or deactivating contours as needed. Active and inactive contours are color-coded (green / red).
Spatial Review (3-Plane View) A sagittal and coronal reconstruction (MPR) is generated from the axial slices, allowing the contour and lesion finding to be checked for spatial consistency.
Complete Technical Analysis The application performs a complete technical analysis of the entire examination (sequence quality, consistency, limitations) and automatically documents it in a Word report. Slice-thickness or slice-spacing issues that cause misalignment between the different sequences – essential for lesion assessment in prostate MRI – are likewise automatically detected and documented.
Lesion Finding Report A second report evaluates any lesions found and provides a comprehensible, reasoned justification in text form – likewise generated automatically.

Storyboard: Lesion Search & MRI Analysis

Schematic sequence showing how a substantiated lesion finding emerges from three image sequences – all illustrations are sketches, not real patient images.

Schematic storyboard – freely created sketch, not real patient images

Principle of the 3-plane view (schematic)

Program Interface (Sample Images)

Sample view of the application using entirely synthetic image data – no real patient reference. Click to enlarge.

Manual contour correction with prostate volume display (synthetic test data)

Rendered using synthetic MR images

3-plane view with diffusion and ADC evaluation (synthetic test data)

Rendered using synthetic MR images

Key Bindings at a Glance

Key	Function
K	Start automatic AI segmentation
F1	Show help
F2	Generate analysis report
F3	View: lesion segment
F5 / Esc	Cancel operation
P	Switch compute mode (GPU / CPU)

Full overview of all key bindings and the operating instructions directly within the application via the Help key.

PI-RADS 2.1

From the Radiological Criteria Discussion to the Algorithm

The lesion search was not handed over as a ready-made formula, but developed through dialogue: first, what radiologically defines a PI-RADS 2.1-suspicious lesion was discussed – only afterward was it worked out together with Claude Code how to translate this description into concrete, computable thresholds. Both are documented here.

💬 Discussed Semantically

The starting point was the radiological PI-RADS 2.1 description of a clinically relevant lesion, with diffusion and ADC as the leading sequences:

Marked diffusion restriction (high signal at high b-value)
Simultaneously, focally reduced ADC values
T2-hypointense area (darker than the surrounding tissue)
Round to oval shape – not diffuse, not linear
Above a clinically plausible minimum size, not just single pixels

The key agreement: none of these features alone is sufficient – only the SIMULTANEOUS presence of all three imaging findings at the same location makes an area suspicious.

🧩 Algorithmically Implemented

Diffusion threshold, adaptive per examination: only the top ~3 % of diffusion values within the prostate contour count as „high“.
ADC window (median-filtered, for lesion detection only – see Technology section): typical restrictive range around 600–900 ×10^-6 mm²/s.
T2 hypointensity: the pixel must lie below the median of the entire slice.
A hard core exists only if all three criteria are met SIMULTANEOUSLY; adjacent pixels that narrowly miss are added with tolerance, isolated regions without a core are discarded.
Shape filter: only round to oval shapes above a minimum size are kept – linear or star-shaped structures (e.g. vessels) are excluded.
3D consistency check across multiple slices and the sagittal/coronal reconstruction, before an area counts as a lesion.
Color coding as a smooth gradient from green → red → orange → yellow → white, ascending with signal intensity within the detected area: white marks the point of highest intensity – where the likelihood of a clinically relevant lesion (PI-RADS 4/5 range) is greatest. A purely visual ranking, not an automatically computed PI-RADS score.

Only the intersection of all three criteria counts as a lesion core – schematic, not a real patient scan