Cursor-rules
PromptBeginner5 minmarkdown
.cursorrules
.llmrules
0
Cursor-rules
{
"model": "DeepCORO_CLIP",
promptBeginner5 min to valuemarkdown
0 views
Jan 15, 2026
Sign in to like and favorite skills
"model": "DeepCORO_CLIP",
Sign in to like and favorite skills
.llmrules
Code Style and Structure
Here are some pieces of information you need to know about this project:
{ "model": "DeepCORO_CLIP", "acceptsMultipleVideos": true, "structure": { "sections": [ { "title": "Dataset", "content": "Trained on over 12 million echocardiography videos paired with text reports from 275,442 studies. Accepts multiple videos from the same exam." }, { "title": "Encoders", "subsections": [ { "title": "Video Encoder", "content": "Utilizes Multiscale Vision Transformer (mVIT) pretrained on the Kinetics dataset. Processes 16-frame clips to capture temporal dynamics, generating a 512-dimensional embedding." }, { "title": "Text Encoder", "content": "Based on BioMedBERT, trained on PubMed abstracts, modified to produce 512-dimensional embeddings." } ] }, { "title": "Contrastive Training", "content": "Aligns video clips and corresponding text reports in a joint embedding space using contrastive loss. Uses batches of 32 video-report pairs." }, { "title": "Multiple Instance Learning (MIL)", "content": "Employs an anatomical attention mechanism to identify the importance of each video and its view for specific anatomical interpretations. Utilizes a view classifier trained on 58 echocardiographic views." }, { "title": "Retrieval-Augmented Interpretation", "content": "Leverages historical report embeddings to retrieve and weight the most relevant reports based on anatomical attention, synthesizing insights from multiple views and videos." }, { "title": "Inference Pipeline", "steps": [ "Categorize videos into views.", "Generate embeddings using the video encoder.", "Apply anatomical attention to aggregate information.", "Produce a comprehensive study-level interpretation." ] }, { "title": "Fine-Tuning", "content": "Pretrained on the entire dataset for 60 epochs. Fine-tuned for 20 epochs on a refined dataset excluding less relevant modalities." }, { "title": "Augmentations", "techniques": [ "RandAugment", "RandomErasing" ] }, { "title": "Evaluation", "content": "Validated through cross-modal retrieval tasks (video-to-text and text-to-video). Outperformed previous models on internal and external test sets." } ] } }