BIMCV-Prostate-Classification

A Repository Containing the code used to classify Clinically Significant Prostate Cancer on BIMCV Prostate MRI Dataset.

Table of Contents

• Introduction
• Dataset
• Models
• Methodology
• Results
• Dependencies
• Installation
• Repository Structure
• Usage
• Interpretability Reports
• References
• Grants and Funding

Introduction

Prostate cancer is a leading cause of cancer-related death among men worldwide. Accurate detection and classification of csPCa using MRI are crucial for diagnosis, treatment planning, and monitoring. This project introduces AI-based models designed to improve the detection of csPCa using MRI.

Dataset

The datasets used in this project include the PI-CAI Challenge dataset [1] and MRI images and clinical data from over 8,000 subjects from the BIMCV Prostate dataset. The dataset is structured using the Medical Imaging Data Structuring (MIDS) framework to ensure robustness and diversity.

Models

3D EfficientNet-B7 (Classification)

A 3D EfficientNet-B7 model is adapted and trained for classifying csPCa from multiparametric MRI scans. The model leverages transfer learning from the PI-CAI dataset to enhance performance. It was adapted using MONAI 1.2.0.

Methodology

1. Data Preprocessing: Image normalization, augmentation, and prostate zone segmentation.
2. Model Training: Utilization of advanced deep learning techniques, including Vision Transformers and Transfer Learning.
3. Model Evaluation: Benchmarking against other models and applying interpretability techniques such as occlusion sensitivity and guided backpropagation.

Results

The 3D EfficientNet-B7 model achieved an AUC of 0.82 in the validation set.

Generated figures and JSON metrics are stored under `outputs/figures/` and `outputs/results/`. Additional comparison plots are under `outputs/figures/results/`.

Dependencies

The current release depends on the following Python libraries:

• monai == 1.2.0
• numpy == 1.23.4
• pandas == 2.1.0
• prettytable == 3.9.0
• ptflops == 0.7
• pygad == 3.2.0
• tensorboard == 2.8.0
• torch == 1.12.1
• torchmetrics == 1.1.2
• torchvision == 0.13.1
• tqdm == 4.62.3

Notebook analysis also uses:

• plotly
• kaleido (static export requires Chrome or kaleido.get_chrome())
• ipywidgets (for tqdm progress bars in notebooks)

Installation

To set up the project environment and code you can install BIMCV_AIKit:

git clone https://github.com/BIMCV-CSUSP/BIMCV-AIKit.git
cd BIMCV-AIKit
pip install -e .

Then, from this repository root, install the local package so modules under src/ are importable:

pip install -e .

Repository Structure

• src/bimcv_prostate/ - Core library
  • data/ - Loaders and transforms
  • models/ - Model definitions
  • training/ - Train/averaging utilities
  • utils/ - Path mapping and config helpers
  • interpretability/ - Shared report generation utilities
  • experiments/clinical/ - Clinical variable experiments
• configs/ - Training configs (JSON)
• scripts/ - CLI entrypoints (train, eval, average_model, interpretability)
• notebooks/ - Analysis notebooks
• outputs/ - Generated artifacts (results, figures, reports)

Usage

1. Data Preparation: Ensure the dataset is available and structured as required. All data preparation and analysis can be found in the Data Structuring and Data Analysis folders.

2. Training: For pre-training with the PI-CAI dataset, download The PI-CAI Challenge: Public Training and Development Dataset. Pretrain using BIMCV-AIKit with:

bimcv_train -c configs/config_picai.json

or

python -m bimcv_aikit.training.train -c configs/config_picai.json

Finally, with the pretrained model, update pretrained_weights_path in configs/config.json and train on BIMCV:

bimcv_train -c configs/config.json

You can also use the helper wrapper:

python scripts/train.py --config configs/config.json

3. Evaluation: Run the main analysis notebook or execute it via the helper script:

python scripts/eval.py --notebook notebooks/Analize_Results.ipynb

4. Model Averaging: Build an averaged backbone from fold checkpoints:

python scripts/average_model.py --fold path/to/fold0.pth --weight 0.25 --fold path/to/fold1.pth --weight 0.25

5. Clinical Variables Experiments: Use the clinical configs under configs/clinical/:

bimcv_train -c configs/clinical/config.json

Interpretability Reports

Two scripts generate per-patient PDF interpretability reports using MONAI visualization methods (Grad-CAM++, Guided Grad-CAM, occlusion sensitivity, and guided backpropagation):

• scripts/interpretability.py - Merges all available segmentations for a session into a single ground-truth mask.
• scripts/interpretability_2.py - Loads all available segmentations and overlays them as multiple contours.

Both scripts now share the same core logic under src/bimcv_prostate/interpretability/report_generator.py to match the project structure and reduce duplication.

Prerequisites

Before running either script, confirm the following:

• You installed the local package with pip install -e . from the repository root.
• ../Files/data_interpretability.csv exists and includes a partition column with a val split.
• The CSV columns t2, adc, and dwi are valid paths after rewrites.
• The XNAT session root exists at ../Files/data/validation/descargas_xnat_segura/.
• The model checkpoints listed inside each script under MODEL_PATHS exist on disk.

Run (Examples)

From the repository root:

python3 scripts/interpretability.py --limit 1 --start-index 0

python3 scripts/interpretability_2.py --limit 1 --start-index 0

Outputs

• Reports are written to ../Results/Patient_Reports_PDF/.
• Filenames follow {ID_XNAT_session}_Report.pdf when that column is available; otherwise they use Patient_{index}.
• If a report already exists, a numeric suffix is appended (for example, _Report_1.pdf).

Configuration Notes

Key settings live near the top of each script in REPORT_CONFIG:

• target_layer controls the Grad-CAM++ layer.
• channel_index selects the background channel shown in the report (for example, ADC).
• class_index selects the class to explain.
• path_rewrites must match the data layout used by ProstateImageDataLoader.

If you want to change shared behavior, prefer editing src/bimcv_prostate/interpretability/report_generator.py rather than duplicating logic inside the scripts.

References

[1] A. Saha, J. S. Bosma, J. J. Twilt, B. van Ginneken, A. Bjartell, A. R. Padhani, D. Bonekamp, G. Villeirs, G. Salomon, G. Giannarini, J. Kalpathy-Cramer, J. Barentsz, K. H. Maier-Hein, M. Rusu, O. Rouvière, R. van den Bergh, V. Panebianco, V. Kasivisvanathan, N. A. Obuchowski, D. Yakar, M. Elschot, J. Veltman, J. J. Fütterer, M. de Rooij, H. Huisman, and the PI-CAI consortium. “Artificial Intelligence and Radiologists in Prostate Cancer Detection on MRI (PI-CAI): An International, Paired, Non-Inferiority, Confirmatory Study”. The Lancet Oncology 2024; 25(7): 879-887.

Grants and Funding

Funded by the Spanish Ministry of Economic Affairs and Digital Transformation (Project MIA.2021.M02.0005 TARTAGLIA, from the Recovery, Resilience, and Transformation Plan financed by the European Union through Next Generation EU funds). TARTAGLIA takes place under the R&D Missions in Artificial Intelligence program, which is part of the Spain Digital 2025 Agenda and the Spanish National Artificial Intelligence Strategy.