Codatta/Refined-TCGA-PRAD-Prostate-Cancer-Pathology-Dataset

Note: Our 245 TCGA cases are ones we identified as having potential for improvement. We plan to upload them in two phases: the first batch of 138 cases, and the second batch of 107 cases in the quality review pipeline, we plan to upload them around early of January, 2025.

Dataset: A Second Opinion on TCGA PRAD Prostate Dataset Labels with ROI-Level Annotations

Overview

This dataset provides enhanced Gleason grading annotations for the TCGA PRAD prostate cancer dataset, supported by Region of Interest (ROI)-level spatial annotations. Developed in collaboration with Codatta and DPath.ai, where DPath.ai launched a dedicated community via Codatta to assemble a network of pathologists, this dataset improved accuracies and granularity of information in the original TCGA-PRAD slide-level labels. The collaborative effort enabled pathologists worldwide to contribute annotations, improving label reliability for AI model training and advancing pathology research. Unlike traditional labeling marketplaces, collaborators a.k.a pathologists retain ownership of the dataset, ensuring their contributions remain recognized, potentially rewarded and valuable within the community. Please cite the dataset in any publication or work to acknowledge the collaborative efforts of Codatta, DPath.ai, and the contributing pathologists.

Motivation

We discovered significant opportunities for data improvement in the diagnostic labels of the TCGA PRAD dataset, where:

• Some labels could be enhanced with additional diagnostic labels a.k.a opinions.
• Some labels lacked granular descriptions of the Gleason patterns.

This presents a challenge for AI pathology models, as reported high accuracy might reflect learning from improved labels. To address this:

• Pathologists re-annotated slides to improve label quality.
• ROI annotations were introduced to clearly differentiate Gleason tumor grades.
• Each annotation is supported by detailed reasoning, providing transparency and justification for corrections.

Dataset Contents

This dataset includes two primary files:

1. Slide-Level Labels (PRAD.csv)

   • Contains comprehensive metadata and diagnostic details:
     • data_profile_uri: A URI that links to Codatta's web application, offering a detailed view of the slide's metadata and its associated data lineage.
     • slide_id: Unique slide identifier.
     • slide_name: TCGA Whole Slide Image (WSI) name.
     • label: Corrected Gleason grade (e.g., 4+3, 5+4).
     • diagnosis: Pathologist-provided reasoning for the label.
     • num_rois: Number of labeled Regions of Interest (ROIs) per slide.

2. ROI-Level Annotations (.geojson)

   • Provides spatial coordinates for tumor regions:
     • Each ROI corresponds to specific Gleason grades (e.g., Grade 3, Grade 4).
     • Compatible with tools like QuPath for interactive visualization.

Key Statistics

Type	Counts
TCGA WSI Total	435
Agree with TCGA's label	190
Refined label and ROI	245

We are currently in the process of uploading data files to meet the quantities mentioned above. This is an ongoing effort to balance community impact and data quality.

Re-Annotation and Labeling Process

1. Curation of Cases for Enhancement: An expert committee identified slides requiring review.
2. Annotation: Junior pathologists performed initial ROI-level annotations.
3. Expert Review: Senior pathologists validated and refined the annotations.
4. Enhancements:
   • Granular ROI labeling for tumor regions.
   • GIntroduction of Minor Grades: For example, Minor Grade 5 indicates <5% of Gleason Grade 5 tumor presence.
   • GPathologist Reasoning: Each label includes a detailed explanation of the annotation process. Some labels can be improved by adding alternative opinions to enhance the labels further.

Improvements Over TCGA Labels

• Accuracy: Enhanced slide-level Gleason labels with additional opinions and improved granularity.
• Granularity: Clear ROI-level annotations for primary, secondary, and minor tumor grades.
• Transparency: Pathologist-provided reasoning ensures a chain of thought for label decisions.

Example:

• Original TCGA label: Gleason Grade 4+3.
• Enhanced label: Gleason Grade 4+3 + Minor Grade 5.

Usage

For AI Training Pipelines

Combine Whole Slide Images (WSI) from TCGA PRAD with this dataset's slide-level labels (PRAD.csv) and ROI annotations (.geojson) to generate high-quality [X, y] pairs.

For Pathology Research

Use the ROI annotations in Whole Slide Images (WSIs) to interactively visualize labeled tumor regions. The slides can be viewed through Codatta's data profile (e.g., https://data.codatta.io/[slide_id]) or other compatible viewers like QuPath. Additionally, explore detailed reasoning behind Gleason grade decisions to gain insights into tumor composition.

How to Load the Dataset

1. CSV File Use pandas to explore slide-level metadata:

import pandas as pd
df = pd.read_csv("PRAD.csv")
print(df.head())

2. GeoJSON File Load ROIs using tools like QuPath or GeoPandas:

import geopandas as gpd
roi_data = gpd.read_file("annotations.geojson")
roi_data.plot()

3. TCGA Whole Slide Images Original WSI files can be downloaded from the GDC Portal. Match WSI filenames with the slide_name column in PRAD.csv for integration.

Example Steps

1. Download TCGA PRAD slides from the GDC Portal.
2. Load PRAD.csv to access corrected labels and reasoning.
3. Visualize ROI annotations using the .geojson files.
4. Train AI models with X = WSI images, y = ROI annotations + slide-level labels.

Licensing

This dataset is licensed under OpenRAIL-M for non-commercial use. Commercial use requires obtaining a separate license. Please contact [email protected] for licensing inquiries.

Credits

This dataset is a collaboration between:

• Codatta: A platform that brings together a community of anonymous pathologists to collaborate on annotation and labeling tasks.
• DPath.ai: The initiator of the project, responsible for defining the task scope, data requirements, quality assurance workflows, and recruiting the initial cohort of expert pathologists.

Contact

For questions, suggestions, or collaborations (launch custom data sourcing and labeling tasks), please reach out via:

• Email: [email protected]
• Website: https://codatta.io