Update README.md

README.md

@@ -1,3 +1,29 @@
----
-license: mit
----
+---
+license: mit
+tags:
+- biology
+- pathology
+- TCGA
+- Prostate
+pretty_name: Refined TCGA-PRAD Prostate Cancer Pathology Dataset
+size_categories:
+- n<1K
+---
+
+Dataset Description: A Second Opinion on TCGA PRAD Prostate Dataset Labels Supported by ROI-Level Annotations 
+Motivation:
+We discovered that some diagnostic labels in the TCGA's PRAD prostate cancer dataset were incorrect, and some are incomplete in terms of describing the Gleason patterns presented on the slides. This presents a significant issue, as many recent AI pathology studies show models achieving higher accuracy on the TCGA dataset. However, if the labels are wrong, a higher accuracy on incorrect labels could suggest that the model is performing worse than it actually is when compared to true labels. To address this, Codatta collaborated with DPath and expert-level pathologists to re-annotate some of the TCGA PRAD dataset, improving the label quality and ensuring more reliable data for AI model training.
+| Type                     | Counts |
+|--------------------------|--------|
+| TCGA WSI Total           | 435    |
+| Agree with TCGA label    | 190    |
+| Disagree with TCGA label | 245    |
+
+Re-Annotation Process:
+To support our new slide-level labels, pathologists meticulously labeled the corresponding regions of interest (ROIs) within each slide. For example, when a slide was labeled as Gleason Grade 4+3, the regions containing Gleason Grade 4 tumor and Gleason Grade 3 tumor were separately marked. This provided a clear indication of the relative proportion of these tumor grades within the slide, helping clarify the grading structure. Each label also includes a paragraph of reasoning, offering a chain of thought that explains how the pathologist arrived at the label, ensuring transparency and justification for the second-opinion annotations.
+
+Improvements Over TCGA Labels:
+In addition to correcting TCGA's slide-level labels, we made the grading system more detailed. The TCGA labels only distinguish between primary and secondary grades for prostate cancer. Our annotations introduce a "Minor" category, which provides a more nuanced understanding of the tumor composition. For example, a slide labeled as Gleason Grade 4+3 might also be annotated with Minor Grade 5, indicating that less than 5% of the slide contains Grade 5 tumor, which is valuable for prognosis and treatment planning.
+
+How to Use:
+This dataset includes two key files to facilitate its use: (1) an Excel file containing comprehensive slide-level information, including the slide name, diagnosis, label, and the number of regions of interest (ROIs) for each slide, and (2) a JSON file that provides the ROI coordinates for each slide. The Excel file serves as a detailed reference for understanding the annotations and slide-level statistics. The JSON file can be directly opened with QuPath or other whole slide image (WSI) viewers, allowing users to visualize the ROIs interactively on the slides. This setup makes it easy to explore the data, validate the annotations, and integrate the dataset into AI training pipelines or pathology workflows.