Source-linked AI summary
Going Deeper through the Gleason Scoring Scale: An Automatic end-to-end System for Histology Prostate Grading and Cribriform Pattern Detection
Julio Silva-Rodríguez, Adrián Colomer, María A. Sales, Rafael Molina, Valery Naranjo
TL;DR
The paper addresses Gleason grading in prostate-biopsy tissue, where grade proportions and severity both matter. It proposes end-to-end CNN-based approaches trained from scratch and reports strong test performance, including cribriform-pattern detection.
Problem
Gleason grades depend not only on tissue proportions but also on the severity of each grade.
Method
The paper proposes end-to-end approaches using a carefully designed CNN trained from scratch to estimate Gleason-grade tissue proportions for combined biopsy scoring.
Results
0.77 Cohen’s quadratic kappa was achieved for test-image patch-level Gleason grading, while cribriform-pattern detection reached 0.82 AUC.
Takeaways & Limitations
The proposed approaches achieved the best results on the authors’ dataset and reached the level of fine-tuned state-of-the-art architectures.
Takeaways & Limitations
The study is limited by intra-observer variability and large heterogeneity within the data, which can affect the figures of merit.
Abstract
from arXiv · showhide
The Gleason scoring system is the primary diagnostic and prognostic tool for prostate cancer. In recent years, with the development of digitisation devices, the use of computer vision techniques for the analysis of biopsies has increased. However, to the best of the authors' knowledge, the development of algorithms to automatically detect individual cribriform patterns belonging to Gleason grade 4 has not yet been studied in the literature. The objective of the work presented in this paper is to develop a deep-learning-based system able to support pathologists in the daily analysis of prostate biopsies. The methodological core of this work is a patch-wise predictive model based on convolutional neural networks able to determine the presence of cancerous patterns. In particular, we train from scratch a simple self-design architecture. The cribriform pattern is detected by retraining the set of filters of the last convolutional layer in the network. From the reconstructed prediction map, we compute the percentage of each Gleason grade in the tissue to feed a multi-layer perceptron which provides a biopsy-level score.mIn our SICAPv2 database, composed of 182 annotated whole slide images, we obtained a Cohen's quadratic kappa of 0.77 in the test set for the patch-level Gleason grading with the proposed architecture trained from scratch. Our results outperform previous ones reported in the literature. Furthermore, this model reaches the level of fine-tuned state-of-the-art architectures in a patient-based four groups cross validation. In the cribriform pattern detection task, we obtained an area under ROC curve of 0.82. Regarding the biopsy Gleason scoring, we achieved a quadratic Cohen's Kappa of 0.81 in the test subset. Shallow CNN architectures trained from scratch outperform current state-of-the-art methods for Gleason grades classification.
1. Introduction
The paper addresses the workload and variability of Gleason grading by developing an end-to-end system for cancer-pattern detection, cribriform detection, grade quantification, and biopsy-level scoring.
- Clinical motivation: Gleason scoring is time-consuming, repetitive, and subject to intra- and inter-pathologist variability.Updated guidelines also recommend reporting the percentage of Gleason grade 4 and the presence of cribriform patterns.
- System objectives: The system detects cancerous regions, identifies cribriform patterns, calculates each Gleason grade’s tissue percentage, and scores the whole biopsy using grade proportion and severity.Its workflow combines patch-level prediction with whole-biopsy scoring.
- Method: The authors develop a patch-level Gleason-grade predictor using a carefully designed CNN architecture trained from scratch.The reconstructed tissue predictions support biopsy-level analysis.
- Results: 0.82 AUC was obtained for cribriform-pattern detection, addressing a previously unstudied detection task.The model retrains filters in the final convolutional layer for cribriform detection.
- Results: 0.77 Cohen’s quadratic kappa was achieved for test-set patch-level Gleason grading, outperforming previous state-of-the-art methods.The paper also reports that the model outperformed fine-tuned state-of-the-art architectures in the authors’ comparison.
2. Materials: SICAP database
SICAPv2 is presented as a large prostate-biopsy WSI resource combining biopsy-level Gleason scores with pixel-level annotations, including cribriform regions, to support end-to-end model development.
- Database contribution: SICAPv2 provides prostate biopsy whole-slide images with biopsy-level labels and pixel-level Gleason-grade annotations.The database was created to address the limited clinical representativeness of small cores and isolated patches.
- Data collection: 155 biopsies from 95 patients were collected, scanned at 40x magnification, and reviewed by expert urogenital pathologists.Combined Gleason scores were assigned per biopsy, with uncertain cases resolved by expert consensus.
- Biopsy labels: The dataset includes 36 non-cancerous biopsies, 40 with Gleason grade 3, 64 with grade 4, and 15 with grade 5.Primary and secondary Gleason-grade co-occurrences are summarized in Figure 4.
- Patch annotations: After filtering, the database contained 4417 non-cancerous patches, 1635 GG3 patches, 3622 GG4 patches, and 665 GG5 patches.Among these, 763 GG4 patches also contained annotated cribriform regions.
3. Methods
The system uses a shallow CNN pipeline for patch-level Gleason grading, cribriform detection, and whole-slide scoring. Patch predictions are reconstructed into tissue-grade percentages that an MLP converts into biopsy-level Gleason scores.
- Patch-Level Gleason Grading: The proposed FSConv model is a self-designed convolutional architecture trained from scratch for patch-level Gleason classification.It uses three convolutional blocks and compares fully connected, global-average-pooling, and global-max-pooling top models.
- Patch-Level Gleason Grading: Global-max pooling is used to reduce sensitivity to the amount and location of cancerous tissue in sliding-window patches.The design addresses cases where patches contain only small portions of cancerous tissue.
- Cribriform Pattern Detection: The cribriform detector fine-tunes the Gleason-grading model by retraining selected convolutional filters and adding a sigmoid output after global-max pooling.Binary cross-entropy is used for this detection task.
- Whole Slide Image Gleason Scoring: Patch-level class probabilities are interpolated into whole-slide probability maps, from which the percentage of each Gleason grade is calculated.Each pixel is assigned a class before grade percentages are computed.
- Whole Slide Image Gleason Scoring: The MLP takes whole-slide Gleason-grade percentages as input and predicts primary and secondary grades through two output branches.Each output uses four target classes: non-cancerous, GG3, GG4, and GG5.
4. Experiments
Experiments evaluate patch-level Gleason grading, cribriform detection, and whole-slide scoring using patient-separated validation procedures and standard classification metrics. The proposed shallow, scratch-trained model achieves strong results across these tasks, including external-database evaluation and biopsy-level scoring.
- FSConv Architecture Benchmarking: 0.7622 accuracy and 0.7125 average F1-score were achieved by FSConv+GMP, surpassing VGG19+GMP+FC at 0.7273 and 0.6772.FSConv+GMP also reached F1-scores of 0.7277 for GG3 and 0.7093 for GG4.
- FSConv Architecture Benchmarking: 2 × 10^7 versus 6 × 10^5 parameters were reported for VGG19+GMP+FC and FSConv+GMP, respectively, favoring FSConv for deployment.The authors also report that scratch-trained self-defined architectures outperform fine-tuned ImageNet architectures for Gleason grading.
- Validation on External Databases: 0.65 macro-averaged F1-score was obtained on the test cohort compared with 0.57 on the Arvaniti et al. database.The model was evaluated on an external database, which the authors identify as a first for patch-level Gleason grading.
- Cribriform Pattern Detection: 0.8225 AUC was achieved for cribriform-pattern detection, with sensitivity 0.8993 and specificity 0.5223 in the reported validation configuration.A slightly higher AUC of 0.8240 was reported for another configuration.
- Whole Slide Image Gleason Scoring: 0.8177 quadratic Cohen’s kappa was obtained for whole-slide Gleason scoring using the proposed MLP.The authors state that the strategy better models the pathologist’s decision to assign a Gleason score to a full slide.
5. Conclusions and future work
The study validates an end-to-end system for patch-level Gleason grading, cribriform detection, tissue quantification, and biopsy-level scoring, while introducing SICAPv2 as a large public annotated database. Its shallow CNN trained from scratch achieved strong reported performance, though annotation variability and grade heterogeneity constrain evaluation.
- The proposed end-to-end system supports pathologists by predicting pixel-level Gleason grades, detecting cribriform patterns, quantifying tissue percentages, and scoring entire biopsies.The pipeline combines local pattern prediction with biopsy-level aggregation through a multi-layer perceptron.
- A concise three-convolutional-layer model trained from scratch achieved the best results on the authors’ dataset, reaching a Cohen’s quadratic kappa of 0.77 in test images.The authors associate this result with a shallow architecture, small receptive field, and reduced parameter count.
- Retraining only the last convolutional-layer filter weights detected cribriform regions in Gleason grade 4 patches with an AUC of 0.82 in the test subset.The work presents this as the first study to contemplate automatic cribriform-pattern detection in prostate histology images.
- The biopsy-level model reached a Cohen’s quadratic kappa of 0.8168 in the test cohort and better reproduced pathologist scoring than approaches based only on the two highest-grade percentages.The score was predicted from the percentages of non-cancerous, Gleason grade 3, 4, and 5 tissues.
- Evaluation is limited by intra-observer annotation variability and substantial heterogeneity within Gleason grades, which complicates balancing patterns across training and testing folds.The authors note that annotation variability may affect the reported figures of merit.
- SICAPv2 contributes a large public database containing pixel-level annotations of prostate biopsies, including Gleason grades and cribriform patterns.Future work includes adding more annotated whole-slide images and jointly predicting additional cancerous patterns.