Spaces:
Running
Running
| from logging import PlaceHolder | |
| import gradio as gr | |
| import os, sys | |
| from npc_bert_models.gradio_demo import * | |
| from npc_bert_models.mlm_module import NpcBertMLM | |
| from npc_bert_models.cls_module import NpcBertCLS | |
| from npc_bert_models.summary_module import NpcBertGPT2 | |
| from npc_bert_models.app_logger import get_logger | |
| import json | |
| class main_window(): | |
| logger = get_logger('main') | |
| def __init__(self): | |
| self.interface = None | |
| self.examples = json.load(open("examples.json", 'r')) | |
| self.logger.info(f"Created {__class__.__name__} instance.") | |
| def initialize(self): | |
| #! Initialize MLM | |
| self.logger.info("Loading MLM interface...") | |
| self.npc_mlm = NpcBertMLM() | |
| self.npc_mlm.load() | |
| with gr.Blocks() as self.mlm_interface: | |
| gr.Markdown("# Masked work prediction\n" | |
| "Enter any sentence. Use the token `[MASK]` to see what the model predicts.\n" | |
| "## Our examples:\n" | |
| "|Input masked sequence|Ground-truth masked word|\n" | |
| "|---------------------|------------------------|\n" | |
| + "\n".join([f"|{a}|{b}|" for a, b in zip(self.examples['mlm-inp'], self.examples['mlm-inp-GT'])])) | |
| with gr.Row(): | |
| with gr.Column(): | |
| inp = gr.Textbox("The tumor is confined in the [MASK].", label='mlm-inp') | |
| btn = gr.Button("Run", variant='primary') | |
| with gr.Column(): | |
| out = gr.Label(num_top_classes=5) | |
| gr.Examples(self.examples['mlm-inp'], inputs=inp, label='mlm-inp') | |
| btn.click(fn=self.npc_mlm.__call__, inputs=inp, outputs=out) | |
| inp.submit(fn=self.npc_mlm.__call__, inputs=inp, outputs=out) | |
| #! Initialize report classification | |
| self.logger.info("Loading BERTCLS interface...") | |
| self.npc_cls = NpcBertCLS() | |
| self.npc_cls.load() | |
| with gr.Blocks() as self.cls_interface: | |
| gr.Markdown(""" | |
| # Report discrimination | |
| In this example we explored how the fine-tuned BERT aids downstream task. We further trained it | |
| to do a simple task of discriminating between reports written for non-NPC patients and NPC patients. | |
| # Disclaimer | |
| The examples are mock reports that is created with reference to authentic reports, they do not | |
| represent any real patients. However, it was written to be an authentic representation of NPC or | |
| patient under investigation for suspected NPC but with negative imaging findings. | |
| """) | |
| with gr.Row(): | |
| with gr.Column(): | |
| inp = gr.TextArea(placeholder="Use examples at the bottom to load example text reports.") | |
| inf = gr.File(file_types=['.txt'], label="Report file (plaintext)", show_label=True, interactive=True) | |
| inf.upload(fn=self._set_report_file_helper, inputs=inf, outputs=inp) | |
| inf.change(fn=self._set_report_file_helper, inputs=inf, outputs=inp) | |
| btn = gr.Button("Run", variant='primary') | |
| with gr.Column(): | |
| out = gr.Label(num_top_classes=2) | |
| # gr.Examples(examples=list(self.examples['reports'].values()), inputs=inp) | |
| gr.Examples(examples="./report_examples", inputs=inf) | |
| btn.click(fn=self.npc_cls.__call__, inputs=inp, outputs=out) | |
| inp.submit(fn=self.npc_cls.__call__, inputs=inp, outputs=out) | |
| #! Initialize report conclusion generation | |
| self.logger.info("Loading Bert-GPT2 interface...") | |
| self.npc_summary = NpcBertGPT2() | |
| self.npc_summary.load() | |
| with gr.Blocks() as self.summary_interface: | |
| gr.Markdown(""" | |
| # Report conclusion generation | |
| In this example we explored how the fine-tunned BERT can aid summarizing the reported items and | |
| generates a conclusion, which includes providing stages of the written reports. | |
| > On this cloud node withonly 2 cpu, it takes ~60 second for this task. | |
| # Disclaimer | |
| Again, similar to the last experiement, the examples we list here are mock reports that are created | |
| with reference to authentic reports and they do not represent any real patients. We essentiall used | |
| the same reports as the last tab, but cropped away the "conclusion" part which we trian the network | |
| to generate. | |
| """) | |
| with gr.Row(): | |
| with gr.Column(): | |
| inp = gr.TextArea(placeholder="Use examples at the bottom to load example text reports.") | |
| inf = gr.File(file_types=['.txt'], label="Report file (plaintext)", show_label=True, interactive=True) | |
| inf.upload(fn=self._set_report_file_helper, inputs=inf, outputs=inp) | |
| inf.change(fn=self._set_report_file_helper, inputs=inf, outputs=inp) | |
| btn = gr.Button("Run", variant='primary') | |
| with gr.Column(): | |
| out = gr.TextArea(placeholder="Conclusion is generated here", interactive=False) | |
| gr.Examples(examples="./report_examples_summary", inputs=inf) | |
| btn.click(fn=self.npc_summary.__call__, inputs=inp, outputs=out) | |
| inp.submit(fn=self.npc_summary.__call__, inputs=inp, outputs=out) | |
| #! Create tab interface | |
| with gr.Blocks() as self.interface: | |
| # Logo | |
| gr.HTML(open("./assets/header.html", 'r').read()) | |
| gr.Markdown(""" | |
| # Introduction | |
| This demo aims to showcase the potential of language models fine-tuned using a meticulously curated dataset of | |
| structured MRI radiology reports for the examination of nasopharyngeal carcinoma (NPC). Our team has a proven | |
| track record in researching the role of AI for the early detection of NPC, having developed an AI system that | |
| achieves high sensitivity and specificity, both exceeding 90%. However, the explainability of the system remains | |
| a significant hurdle for clinical application. This challenge is not unique to our system but is pervasive in the | |
| development of AI for radiology. Therefore, in this pilot study, we investigate the capacity of language models | |
| to comprehend the context of the disease. Our aim is to explore the integration of language models into our | |
| existing system to enhance its explainability. | |
| # Affliations | |
| * Dr. M.Lun Wong, Dept. Imaging and Interventional Radiology. The Chinese University of Hong Kong | |
| # Disclaimer | |
| This software is provided as is and it is not a clinically validated software. The authors disclaim any responsibility | |
| arising as a consequence from using this demo. | |
| """) | |
| tabs = gr.TabbedInterface([self.mlm_interface, self.cls_interface, self.summary_interface], | |
| tab_names=["Masked Language Model", "Report classification", "Report conclusion generation"]) | |
| def lauch(self): | |
| self.interface.launch(allowed_paths=['assets']) | |
| pass | |
| def _set_report_file_helper(self, file_in): | |
| try: | |
| text = open(file_in, 'r').read() | |
| return text | |
| except: | |
| print(f"Cannot read file {file_in}") | |
| # Do nothing | |
| pass | |
| if __name__ == '__main__': | |
| mw = main_window() | |
| mw.initialize() | |
| mw.lauch() | |