PANDA Prostate Cancer Classification Dataset
| Modality | Data Format | Publisher | Licence |
| Light Microscope | TIFF | PANDA Challenge | Non-Commercial |
Prostate cancer is a global health concern with over 1 million new diagnoses annually, resulting in more than 350,000 deaths each year. Early and accurate diagnosis is key to reducing mortality rates. The Gleason grading system, the primary method for diagnosing and estimating the severity of prostate cancer, has significant limitations, including inter-observer variability among pathologists. This variability can lead to both underdiagnosis and overdiagnosis, with the former risking missed severe cases and the latter leading to unnecessary treatment.
The grading process identifies cancer tissue and classifies it into Gleason patterns of 3, 4, or 5 based on the tumor’s architectural growth patterns. These Gleason scores are then translated into ISUP grades that play a vital role in treatment decisions. Given these challenges and the critical importance of accurate grading, the PANDA dataset provides an invaluable resource for developing machine learning models that could offer more consistent and precise grading, thereby improving patient outcomes.
About Prostate Cancer Classification Dataset
The PANDA Prostate Cancer Classification Dataset is an extensive collection aimed at advancing the diagnostics of prostate cancer (PCa), the second most common cancer among men worldwide. Comprising 21,135 files with a total size of 411.9 GB, the dataset includes high-resolution microscopy scans of prostate biopsy samples in TIFF format, alongside metadata in CSV files.
These scans are intended to facilitate the development of machine learning models that can classify the severity of prostate cancer based on the Gleason grading system, subsequently converting these into ISUP grades ranging from 1-5. This grading system is critical for determining the appropriate treatment for patients.
However, it is important to note that the dataset comes with the challenge of imperfect labels, reflecting the real-world difficulty even experienced pathologists face when interpreting these slides. The imperfections in labeling add a layer of complexity to model training but also offer the potential for significant medical value if a consistent and accurate model can be developed.