
Research: Medical AI
[Analyze pathology images (bladder cancer)]
REG 2024(REport Generation for Pathology using Giga-pixel Whole Slide Images in Bladder Tumor, 2024)
Research Backgroud
The significant advancements in vision-language foundation models have enabled a wide range of innovative applications in the medical domain. One notable example is Image Captioning, a technology that utilizes image understanding to generate natural language descriptions.
​
In the context of pathology, applying Image Captioning to giga-pixel pathology images poses a substantial technical challenge. These images require advanced analytical techniques, such as slide-level feature extraction, which involves high-level computational processes to extract meaningful features from large and complex medical images. Tackling these challenges has been the focus of several recent studies, aiming to harness natural language generation (NLG) for automated pathology report creation.
​
The development of an automated pathology report generation model using Image Captioning is expected to significantly advance pathology image analysis. It offers the potential to achieve a higher level of technical sophistication while addressing critical issues such as the global shortage of pathologists. By enhancing diagnostic productivity and efficiency, this approach could make a transformative impact on the field of pathology.
Research Data (PIT_BLT: Pathology Image and Text for Bladder Tumor)
The data used in this study comprises bladder cancer pathology image data, a subset of urinary system cancer pathology data collected by Korea University Medical Center as part of the 2023 AI Training Data Construction Project. The pathology reports were reconstructed based on the CAP guidelines (Urinary bladder, Resection, v4.2.0.0) by three pathology specialists. Any discrepancies in interpretation were resolved through discussion and consensus.
The reports include the following key elements: Procedure, Histologic Type, Histologic Grade, Tumor Extent, and Muscularis Propria. The Histologic Type was documented according to the WHO classification (5th edition), and for histologic subtypes, only the presence or absence of squamous differentiation was recorded. Additionally, cases with granulomatous inflammation were noted as part of other findings.
Data Set: PIT_BLT (Pathology Image and Text for Bladder Tumor)
-
Composition: Pairs of WSI (Whole Slide Image) pathology images and corresponding pathology reports
-
Image (WSI): x200 magnification, 0.5 µm/pixel resolution, stored in .tiff format (converted from its original formats)
-
Text (Label): Pathology reports, stored in .txt format