https://atlas.rsna.org/schemas/2025-11/model.json

Prostate158 - An expert-annotated 3T MRI dataset and algorithm for prostate cancer detection

https://doi.org/10.1016/j.compbiomed.2022.105817

lisa.christine.adams@gmail.com

LCA is grateful for her participation in the BIH Charité – Junior Clinician and Clinician Scientist Program and KKB is grateful for his participation in the BIH Charité Digital Clinician Scientist Program, all funded by the Charité – Universitätsmedizin Berlin and the Berlin Institute of Health.

All contributions to this study were approved by the institutional ethics committee, including a waiver of informed consent, and performed in accordance with data protection policy.

This study presents trained models for automatic segmentation of anatomical zones and areas of suspected prostate cancer in biparametric 3.0 T MRI. The training dataset, training code, and model weights are publicly available.

U-ResNet consisting of six down-sampling blocks and five up-sampling blocks. Each down-sampling block had four convolutional layers, batch normalization, dropout (15%), and PReLU activation. Up-sampling blocks used transposed convolution and convolutional layers with batch normalization, dropout, and PReLU. Trained for 500 epochs with Novograd optimizer, batch size of two, weight decay of 0.01, and one-cycle learning rate scheduler (initial 0.001). Loss function combined soft Dice similarity coefficient and cross-entropy.

The models are publicly available to other researchers along with all segmentation masks at https://github.com/kbressem/prostate158. The test set is not available for download to prevent overfitting, but trained models can be evaluated on the test set using the Grand Challenge platform (https://prostate158.grand-challenge.org).

Improved lesion detection, targeted biopsy planning, radiotherapy planning, improved volume estimation for disease progression assessment, computational detection of prostate zones and prostate cancer.

Diagnosis

Fully automated deep learning pipeline.

Segmentation of anatomical zones (central gland, peripheral zone) and suspicious lesions for prostate cancer (PI-RADS score ≥ 4) in biparametric 3.0 T MRI.

Biparametric 3T prostate MRIs, specifically T2-weighted (T2w) sequences and diffusion-weighted imaging (DWI) with apparent diffusion coefficient (ADC) maps. For anatomical zones, axial T2w sequences were used. For tumor segmentation, T2w, ADC, and DWI sequences were used as concatenated input.

The code to reproduce the training can be found at https://github.com/kbressem/prostate158. Images were pre-processed using intensity normalization (N4ITK), resampling to isotropic voxel spacing of 0.5 × 0.5 × 0.5 mm, and center cropping (removing 20% of image margins). During inference, batch size was reduced to one, and sliding window inference (96 × 96 x 96 voxels with 50% overlap) was used. Prediction maps were one-hot encoded, keeping only the largest connected component for each class.

Comparatively small sample size (158 patients). Lack of dynamic contrast-enhanced (DCE) sequences. Only PI-RADS 4 and 5 lesions were segmented as cancerous lesions, potentially limiting comparability with studies including PI-RADS 3. MRI scanners were from the same vendor (Siemens) and all MRIs had the same in-plane resolution, which could limit generalizability to datasets from different manufacturers. The model was largely trained with segmentations from a single rater, potentially reflecting personal biases and affecting performance against other raters.

The study openly published all training data, all code, and all model weights, making the research fully reproducible.