some updates

app.py

@@ -3,31 +3,25 @@ import text2image
 import image2image
 import visualization
 import streamlit as st
-import streamlit.components.v1 as components
-
-
+import details
 
 st.set_page_config(layout="wide")
 
 st.sidebar.title("WebPLIP")
-
-components.html('''
-    <!-- Google tag (gtag.js) -->
-    <script async src="https://www.googletagmanager.com/gtag/js?id=G-KPF04V95FN"></script>
-    <script>
-    window.dataLayer = window.dataLayer || [];
-    function gtag(){dataLayer.push(arguments);}
-    gtag('js', new Date());
-    gtag('config', 'G-KPF04V95FN');
-    </script>
-''')
+st.sidebar.markdown("## Menu")
 
 PAGES = {
     "Introduction": home,
+    "Details": details,
     "Text to Image": text2image,
     "Image to Image": image2image,
     "Visualization": visualization,
 }
 
 page = st.sidebar.radio("", list(PAGES.keys()))
+st.sidebar.markdown("## Links")
+
+st.sidebar.markdown("[PLIP Model](https://huggingface.co/vinid/plip)")
+st.sidebar.markdown("[OpenPath Dataset]()")
+st.sidebar.markdown("[PLIP Code](https://github.com/vinid/path_eval)")
 PAGES[page].app()

details.py

@@ -0,0 +1,61 @@
+from pathlib import Path
+import streamlit as st
+import streamlit.components.v1 as components
+from PIL import Image
+
+def read_markdown_file(markdown_file):
+    return Path(markdown_file).read_text()
+
+
+def app():
+    #intro_markdown = read_markdown_file("introduction.md")
+    #st.markdown(intro_markdown, unsafe_allow_html=True)
+    st.markdown("# Leveraging medical Twitter to build a visual-language foundation model for pathology")
+
+    st.markdown("The lack of annotated publicly available medical images is a major barrier for innovations. At the same time, many de-identified images and much knowledge are shared by clinicians on public forums such as medical Twitter. Here we harness these crowd platforms to curate OpenPath, a large dataset of <b>208,414</b> pathology images paired with natural language descriptions. This is the largest public dataset for pathology images annotated with natural text. We demonstrate the value of this resource by developing PLIP, a multimodal AI with both image and text understanding, which is trained on OpenPath. PLIP achieves state-of-the-art zero-shot and few-short performance for classifying new pathology images across diverse tasks. Moreover, PLIP enables users to retrieve similar cases by either image or natural language search, greatly facilitating knowledge sharing. Our approach demonstrates that publicly shared medical data is a tremendous opportunity that can be harnessed to advance biomedical AI.", unsafe_allow_html=True)
+
+    fig1ab = Image.open('resources/4x/Fig1ab.png')
+    st.image(fig1ab, caption='OpenPath Dataset', output_format='png')
+    st.caption('An example of tweet')
+    components.html('''
+        <blockquote class="twitter-tweet">
+            <a href="https://twitter.com/xxx/status/1580753362059788288"></a>
+        </blockquote>
+        <script async src="https://platform.twitter.com/widgets.js" charset="utf-8">
+        </script>
+        ''',
+    height=500)
+
+
+    st.markdown("#### PLIP is trained on the largest public vision–language pathology dataset: OpenPath")
+
+    col1, col2 = st.columns([1, 1])
+    with col1:
+        st.markdown("Following the usage policy and guidelines from Twitter and other entities, we established so far the largest public vision–language pathology dataset. To ensure the quality of the data, OpenPath followed rigorous protocols for cohort inclusion and exclusion, including the removal of retweets, sensitive tweets, and non-pathology images, as well as text cleaning.", unsafe_allow_html=True)
+        st.markdown("The final OpenPath dataset consists of:", unsafe_allow_html=True)
+        st.markdown("- Tweets: 116,504 image–text pairs from Twitter posts (tweets) during Mar. 21, 2006 – Nov. 15, 2022 across 32 pathology subspecialty-specific hashtags;", unsafe_allow_html=True)
+        st.markdown("- Replies: 59,869 image–text pairs from the associated replies that received the highest number of likes in the tweet, if applicable;", unsafe_allow_html=True)
+        st.markdown("- PathLAION: 32,041 additional image–text pairs from the Internet which are outside from the Twitter community extracted from the LAION dataset.", unsafe_allow_html=True)
+        st.markdown("Leveraging the largest publicly available pathology dataset which contains image–text pairs across 32 different pathology subspecialty-specific hashtags, where each image has detailed text descriptions, we fine-tuned a pre-trained CLIP model and proposed a multimodal deep learning model for pathology, PLIP.", unsafe_allow_html=True)
+    with col2:
+        fig1c = Image.open('resources/4x/Fig1c.png')
+        st.image(fig1c, caption='Pathology hashtags in Twitter', output_format='png')
+        fig1d = Image.open('resources/4x/Fig1d.png')
+        st.image(fig1d, caption='Number of words in sentence', output_format='png')
+
+
+
+    st.markdown("#### PLIP is trained with connecting the image and text via contrastive learning")
+
+    col1, col2 = st.columns([3, 1])
+    with col1:
+        st.markdown("The proposed PLIP model generates two embedding vectors from both the text and image encoders. These vectors were then forced to be similar for each of the paired image and text vectors and dissimilar for non-paired image and text pairs via contrastive learning.", unsafe_allow_html=True)
+        fig1e = Image.open('resources/4x/Fig1e.png')
+        st.image(fig1e, caption='PLIP training', output_format='png')
+
+    with col2:
+        fig1f = Image.open('resources/4x/Fig1f.png')
+        st.image(fig1f, caption='Training illustration', output_format='png')
+
+
+

home.py

@@ -1,9 +1,6 @@
 from pathlib import Path
 import streamlit as st
 import streamlit.components.v1 as components
-import plotly.figure_factory as ff
-import numpy as np
-import pandas as pd
 from PIL import Image
 
 def read_markdown_file(markdown_file):
@@ -11,59 +8,34 @@ def read_markdown_file(markdown_file):
 
 
 def app():
-    #intro_markdown = read_markdown_file("introduction.md")
-    #st.markdown(intro_markdown, unsafe_allow_html=True)
-    st.markdown("# Leveraging medical Twitter to build a visual-language foundation model for pathology")
 
-    col1, col2 = st.columns([2, 1])
-    with col1:
-        st.markdown("The lack of annotated publicly available medical images is a major barrier for innovations. At the same time, many de-identified images and much knowledge are shared by clinicians on public forums such as medical Twitter. Here we harness these crowd platforms to curate OpenPath, a large dataset of <b>208,414</b> pathology images paired with natural language descriptions. This is the largest public dataset for pathology images annotated with natural text. We demonstrate the value of this resource by developing PLIP, a multimodal AI with both image and text understanding, which is trained on OpenPath. PLIP achieves state-of-the-art zero-shot and few-short performance for classifying new pathology images across diverse tasks. Moreover, PLIP enables users to retrieve similar cases by either image or natural language search, greatly facilitating knowledge sharing. Our approach demonstrates that publicly shared medical data is a tremendous opportunity that can be harnessed to advance biomedical AI.", unsafe_allow_html=True)
-        
-        fig1ab = Image.open('resources/4x/Fig1ab.png')
-        st.image(fig1ab, caption='OpenPath Dataset', output_format='png')
-    with col2:
-        st.caption('An example of tweet')
-        components.html('''
-            <blockquote class="twitter-tweet">
-                <a href="https://twitter.com/xxx/status/1580753362059788288"></a>
-            </blockquote>
-            <script async src="https://platform.twitter.com/widgets.js" charset="utf-8">
-            </script>
-            ''',
-        height=500)
+    st.markdown("# A visual-language foundation model for pathology")
+    st.markdown("This is a webapp for PLIP, our new fundational AI model for pathology and OpenPath our new dataset, **from our recent work**: Leveraging medical Twitter to build a visual-language foundation model for pathology")
+    st.markdown("### Pathology Language and Image Pretraining (PLIP)\n We develop PLIP, a multimodal AI with both image and text understanding. PLIP achieves state-of-the-art zero-shot and few-short performance for classifying new pathology images across diverse tasks. Moreover, PLIP enables users to retrieve similar cases by either image or natural language search, greatly facilitating knowledge sharing. Our approach demonstrates that publicly shared medical data is a tremendous opportunity that can be harnessed to advance biomedical AI.")
 
+    fig1e = Image.open('resources/4x/Fig1e.png')
+    st.image(fig1e, caption='PLIP training procedure', output_format='png')
 
-    st.markdown("#### PLIP is trained on the largest public vision–language pathology dataset: OpenPath")
+    st.markdown("### OpenPath Dataset\nThe lack of annotated publicly available medical images is a major barrier for innovations. At the same time, many de-identified images and much knowledge are shared by clinicians on public forums such as medical Twitter. Here we harness these crowd platforms to curate OpenPath, a large dataset of **208,414** pathology images paired with natural language descriptions")
 
-    col1, col2 = st.columns([1, 1])
-    with col1:
-        st.markdown("Following the usage policy and guidelines from Twitter and other entities, we established so far the largest public vision–language pathology dataset. To ensure the quality of the data, OpenPath followed rigorous protocols for cohort inclusion and exclusion, including the removal of retweets, sensitive tweets, and non-pathology images, as well as text cleaning.", unsafe_allow_html=True)
-        st.markdown("The final OpenPath dataset consists of:", unsafe_allow_html=True)
-        st.markdown("- Tweets: 116,504 image–text pairs from Twitter posts (tweets) during Mar. 21, 2006 – Nov. 15, 2022 across 32 pathology subspecialty-specific hashtags;", unsafe_allow_html=True)
-        st.markdown("- Replies: 59,869 image–text pairs from the associated replies that received the highest number of likes in the tweet, if applicable;", unsafe_allow_html=True)
-        st.markdown("- PathLAION: 32,041 additional image–text pairs from the Internet which are outside from the Twitter community extracted from the LAION dataset.", unsafe_allow_html=True)
-        st.markdown("Leveraging the largest publicly available pathology dataset which contains image–text pairs across 32 different pathology subspecialty-specific hashtags, where each image has detailed text descriptions, we fine-tuned a pre-trained CLIP model and proposed a multimodal deep learning model for pathology, PLIP.", unsafe_allow_html=True)
-    with col2:
-        fig1c = Image.open('resources/4x/Fig1c.png')
-        st.image(fig1c, caption='Pathology hashtags in Twitter', output_format='png')
-        fig1d = Image.open('resources/4x/Fig1d.png')
-        st.image(fig1d, caption='Number of words in sentence', output_format='png')
+    fig1ab = Image.open('resources/4x/Fig1ab.png')
+    st.image(fig1ab, caption='OpenPath Dataset', output_format='png')
 
 
+    st.markdown("### Documentation\n"
+                "This webapp comes with different functionalities.\n"
+                "* Details: The details page guides you through our work.\n"
+                "* Text to Image: allows users to perform text search on a database of images.\n"
+                "* Image to Image: allows users to perform image search on a database of images.\n"
+                "")
 
-    st.markdown("#### PLIP is trained with connecting the image and text via contrastive learning")
-
-    col1, col2 = st.columns([3, 1])
-    with col1:
-        st.markdown("The proposed PLIP model generates two embedding vectors from both the text and image encoders. These vectors were then forced to be similar for each of the paired image and text vectors and dissimilar for non-paired image and text pairs via contrastive learning.", unsafe_allow_html=True)
-        fig1e = Image.open('resources/4x/Fig1e.png')
-        st.image(fig1e, caption='PLIP training', output_format='png')
-
-    with col2:
-        fig1f = Image.open('resources/4x/Fig1f.png')
-        st.image(fig1f, caption='Training illustration', output_format='png')
-
+    st.markdown("### Other Links\n"
+                "* Download OpenPath\n"
+                "* Code to reproduce [PLIP ](https://github.com/vinid/path_eval) results\n"
+                "* Link to the [PLIP Model](https://huggingface.co/vinid/plip)\n"
+                "")
 
+    st.markdown("### Important Information\n")
 
     st.markdown('Disclaimer')
     st.caption('Please be advised that this function has been developed in compliance with the Twitter policy of data usage and sharing. It is important to note that the results obtained from this function are not intended to constitute medical advice or replace consultation with a qualified medical professional. The use of this function is solely at your own risk and should be consistent with applicable laws, regulations, and ethical considerations. We do not warrant or guarantee the accuracy, completeness, suitability, or usefulness of this function for any particular purpose, and we hereby disclaim any liability arising from any reliance placed on this function or any results obtained from its use. If you wish to review the original Twitter post, you should access the source page directly on Twitter.')

introduction.md

@@ -1,6 +0,0 @@
-
-# Leveraging medical Twitter to build a visual-language foundation model for pathology
-
-The lack of annotated publicly available medical images is a major barrier for innovations. At the same time, many de-identified images and much knowledge are shared by clinicians on public forums such as medical Twitter. Here we harness these crowd platforms to curate OpenPath, a large dataset of 208,414 pathology images paired with natural language descriptions. This is the largest public dataset for pathology images annotated with natural text. We demonstrate the value of this resource by developing PLIP, a multimodal AI with both image and text understanding, which is trained on OpenPath. PLIP achieves state-of-the-art zero-shot and few-short performance for classifying new pathology images across diverse tasks. Moreover, PLIP enables users to retrieve similar cases by either image or natural language search, greatly facilitating knowledge sharing. Our approach demonstrates that publicly shared medical data is a tremendous opportunity that can be harnessed to advance biomedical AI.
-
-![Alt Text](resources/4x/Fig1.png)

requirements.txt

@@ -1,7 +1,7 @@
 git+https://github.com/openai/CLIP.git
 torch
 transformers
-pandas
+pandas==1.4.1
 numpy
 Pillow
 streamlit==1.19.0

text2image.py

@@ -2,15 +2,12 @@ import streamlit as st
 import pandas as pd
 import numpy as np
 from PIL import Image
-import requests
 import pickle
 import tokenizers
-from io import BytesIO
 import torch
+
 from transformers import (
-    VisionTextDualEncoderModel,
-    AutoFeatureExtractor,
-    AutoTokenizer,
+
     CLIPModel,
     AutoProcessor
 )
@@ -63,7 +60,7 @@ def embed_texts(model, texts, processor):
 def app():
 
     st.title('Text to Image Retrieval')
-    st.markdown('#### A pathology image search engine that correlate texts directly with images.')
+    st.markdown('#### A pathology image search engine that geos from texts to images.')
     
     col1, col2 = st.columns([1,1])
     with col1:
@@ -76,33 +73,33 @@ def app():
     meta, image_embedding, text_embedding, validation_subset_index = init()
     model, processor = load_path_clip()
 
-    st.markdown('#### Demo')
+    st.markdown('### Search')
+    st.markdown('How to use this: first of all, select a dataset on which to do retrieval.\n'
+                'Then, either select a predefined search query or input one yourself.')
+
 
     col1, col2 = st.columns(2)
     with col1:
-        data_options = ["All twitter data (03/21/2006 — 01/15/2023)",
-                        "Twitter validation data (11/16/2022 — 01/15/2023)"]
-        st.radio(
-            "Choose dataset for image retrieval 👉",
+        data_options = ["All Twitter Data (03/21/2006 — 01/15/2023)",
+                        "Validation Twitter data (11/16/2022 — 01/15/2023)"]
+        st.selectbox(
+            "Dataset",
             key="datapool",
             options=data_options,
         )
+
     with col2:
         retrieval_options = ["Image only",
-                            "Text and image (beta)",
+                             "Text and image (beta)",
                              ]
         st.radio(
-            "Similarity calcuation Mapping input with 👉",
+            "Similarity calcuation 👉",
             key="calculation_option",
             options=retrieval_options,
         )
 
-        
-
     col1, col2 = st.columns(2)
-    #query = st.text_input('Search Query', '')
-    col1_submit = False
-    show = False
+
     with col1:
         # Create selectbox
         examples = ['Breast tumor surrounded by fat',
@@ -112,42 +109,36 @@ def app():
                     'Breast cancer DCIS',
                     'Papillary carcinoma in breast tissue',
                     ]
-        query_1 = st.selectbox("Please select an example query", options=examples)
-        #st.info(f":white_check_mark: The written option is {query_1} ")
+        query_1 = st.selectbox("Select an example", options=examples)
+
         col1_submit = True
-        show = True
-        
+
     with col2:
         form = st.form(key='my_form')
         query_2 = form.text_input(label='Or input your custom query:')
         submit_button = form.form_submit_button(label='Submit')
-    
+
     if submit_button:
         col1_submit = False
-        show = True
-
 
     if col1_submit:
         query = query_1
     else:
         query = query_2
 
-
     input_text = embed_texts(model, [query], processor)[0].detach().cpu().numpy()
     input_text = input_text/np.linalg.norm(input_text)
-    
-    
-    if st.session_state.calculation_option == retrieval_options[0]: # Image only
+
+    # Sort IDs by cosine-similarity from high to low
+
+    if st.session_state.calculation_option == retrieval_options[0]:  # Image only
         similarity_scores = input_text.dot(image_embedding.T)
-    else: # Text and Image
+    else:  # Text and Image
         similarity_scores_i = input_text.dot(image_embedding.T)
         similarity_scores_t = input_text.dot(text_embedding.T)
-        similarity_scores_i = similarity_scores_i/np.max(similarity_scores_i)
-        similarity_scores_t = similarity_scores_t/np.max(similarity_scores_t)
-        similarity_scores = (similarity_scores_i + similarity_scores_t)/2
-
-
-
+        similarity_scores_i = similarity_scores_i / np.max(similarity_scores_i)
+        similarity_scores_t = similarity_scores_t / np.max(similarity_scores_t)
+        similarity_scores = (similarity_scores_i + similarity_scores_t) / 2
 
     ############################################################
     # Get top results
@@ -166,7 +157,7 @@ def app():
     ############################################################
     # Display results
     ############################################################
-    
+
     st.markdown('Your input query: %s' % query)
     st.markdown('#### Top 5 results:')
     topk_options = ['1st', '2nd', '3rd', '4th', '5th']
@@ -203,14 +194,6 @@ def app():
 
 
 
-
-
-
-
-
-
-
-
     st.markdown('Disclaimer')
     st.caption('Please be advised that this function has been developed in compliance with the Twitter policy of data usage and sharing. It is important to note that the results obtained from this function are not intended to constitute medical advice or replace consultation with a qualified medical professional. The use of this function is solely at your own risk and should be consistent with applicable laws, regulations, and ethical considerations. We do not warrant or guarantee the accuracy, completeness, suitability, or usefulness of this function for any particular purpose, and we hereby disclaim any liability arising from any reliance placed on this function or any results obtained from its use. If you wish to review the original Twitter post, you should access the source page directly on Twitter.')