**PATHALIGN: A Vision-Language Model for Whole Slide Images in Histopathology** This paper presents PATHALIGN, a vision-language model designed to process whole slide images (WSIs) in histopathology, leveraging curated text from pathology reports. The model aims to address the challenges of analyzing gigapixel-scale WSIs and integrating them with text descriptions for improved diagnostic accuracy and efficiency. **Key Contributions:** 1. **Model Architecture:** PATHALIGN is based on the BLIP-2 framework, using a patch-level encoder and a large language model (LLM) for text generation. 2. **Data:** The model is trained on a de-identified dataset of over 350,000 WSIs and diagnostic text pairs, covering a wide range of diagnoses, procedure types, and tissue types. 3. **Evaluation:** Pathologist evaluations show that the model-generated text is rated as accurate, with clinically significant errors or omissions rated at 78% accuracy on average. 4. **Applications:** PATHALIGN enables applications such as text retrieval, image-to-text retrieval, WSI classification, and case prioritization, demonstrating its potential in digital histopathology. **Methods:** 1. **Data Curation:** The dataset is curated to pair WSIs with diagnostic text from pathology reports, addressing the challenge of aligning WSIs with specific portions of reports. 2. **Model Training:** PATHALIGN is trained using a two-stage approach, first aligning the WSI and text encoders, and then integrating a frozen LLM for text generation. 3. **Evaluation:** Pathologist evaluations and automatic metrics are used to assess the model's performance in text retrieval, text generation, and WSI classification. **Results:** 1. **Text Retrieval and Generation:** Pathologist evaluations show high accuracy in text retrieval and generation, with generated text often preferred over original diagnostic text. 2. **WSI Classification:** PATHALIGN performs well on tasks such as subtyping of non-small cell lung cancer, renal cell carcinoma, and breast cancer, as well as procedure type classification. 3. **Case Prioritization:** PATHALIGN demonstrates the ability to prioritize cases based on severity, highlighting its potential for clinical applications. **Discussion:** The paper discusses the limitations and future directions, including the need for larger datasets and more sophisticated data curation methods. PATHALIGN shows promise in aligning WSIs with text descriptions, enabling advanced applications in histopathology. **Conclusion:** PATHALIGN represents a significant advancement in the field of computational pathology, providing a robust framework for analyzing WSIs and integrating text descriptions to enhance diagnostic accuracy and efficiency.**PATHALIGN: A Vision-Language Model for Whole Slide Images in Histopathology** This paper presents PATHALIGN, a vision-language model designed to process whole slide images (WSIs) in histopathology, leveraging curated text from pathology reports. The model aims to address the challenges of analyzing gigapixel-scale WSIs and integrating them with text descriptions for improved diagnostic accuracy and efficiency. **Key Contributions:** 1. **Model Architecture:** PATHALIGN is based on the BLIP-2 framework, using a patch-level encoder and a large language model (LLM) for text generation. 2. **Data:** The model is trained on a de-identified dataset of over 350,000 WSIs and diagnostic text pairs, covering a wide range of diagnoses, procedure types, and tissue types. 3. **Evaluation:** Pathologist evaluations show that the model-generated text is rated as accurate, with clinically significant errors or omissions rated at 78% accuracy on average. 4. **Applications:** PATHALIGN enables applications such as text retrieval, image-to-text retrieval, WSI classification, and case prioritization, demonstrating its potential in digital histopathology. **Methods:** 1. **Data Curation:** The dataset is curated to pair WSIs with diagnostic text from pathology reports, addressing the challenge of aligning WSIs with specific portions of reports. 2. **Model Training:** PATHALIGN is trained using a two-stage approach, first aligning the WSI and text encoders, and then integrating a frozen LLM for text generation. 3. **Evaluation:** Pathologist evaluations and automatic metrics are used to assess the model's performance in text retrieval, text generation, and WSI classification. **Results:** 1. **Text Retrieval and Generation:** Pathologist evaluations show high accuracy in text retrieval and generation, with generated text often preferred over original diagnostic text. 2. **WSI Classification:** PATHALIGN performs well on tasks such as subtyping of non-small cell lung cancer, renal cell carcinoma, and breast cancer, as well as procedure type classification. 3. **Case Prioritization:** PATHALIGN demonstrates the ability to prioritize cases based on severity, highlighting its potential for clinical applications. **Discussion:** The paper discusses the limitations and future directions, including the need for larger datasets and more sophisticated data curation methods. PATHALIGN shows promise in aligning WSIs with text descriptions, enabling advanced applications in histopathology. **Conclusion:** PATHALIGN represents a significant advancement in the field of computational pathology, providing a robust framework for analyzing WSIs and integrating text descriptions to enhance diagnostic accuracy and efficiency.