303-TITAN: Multimodal Whole-Slide Pathology Foundation Model
The paper describes the development and extensive evaluation of the TITAN (Transformer-based pathology Image and Text Alignment Network), a new multimodal foundation model for computational pathology. TITAN is pretrained on a massive dataset of over 335,000 whole-slide images (WSIs) using a combination of vision-only learning and subsequent vision-language alignment with both synthetic fine-grained captions and clinical pathology reports. The model utilizes a Vision Transformer architecture, enhanced with ALiBi positional encoding, to manage the immense scale and complexity inherent in…
The skinny
The skinny isn't ready yet — notes appear once the transcript is processed.